AI coding assistants are worth a closer look. For a long time we treated large language models (LLMs) like translation layers: type a thought, get code back. But if you have spent any time debugging production issues with these tools, you have probably watched them “hallucinate” in ways that make no sense. A recent investigation by Shuyang turned up an odd one: a coding assistant that started replying in Korean when prompted in Chinese, because of a few technical English tokens like run.py.
That is not just a quirky bug. It shows how the AI Embedding Space really works. I have been wrestling with WordPress since the 4.x days and have seen my share of “ghosts,” and this one says a lot about where our toolchain is heading.
The engineering attractor field
Most developers assume an LLM organizes data by language, with English, Chinese, and Korean in separate areas. The AI Embedding Space is instead structured by task registers, and there is a specific region of the vector space set aside for “Engineering/Code.”
When you type a prompt like “Please review this branch,” you are not just sending words, you are moving a vector toward an “attractor basin.” Chinese, despite its huge number of speakers, tends to have a thinner representation in high-level engineering corpora than English. So when you mix Chinese technical queries with English tokens (like commit, PR, or diff), the model drifts out of the “Chinese” region and into the “Engineering” one.
So why Korean? In Shuyang’s experiment, as the English words took over, the “Korean similarity” score spiked before the English one won out. That looks like a phase transition. The model lands in a linguistic “no-man’s-land” and grabs the nearest technical register, which in some models happens to be a Korean-English hybrid cluster.
Picturing the drift
Think of it like sprinting a marathon. You start in one lane (Chinese), but the faster you go by adding technical tokens, the more you drift toward the center of the track (English engineering). Lose your footing and you can end up in the wrong lane. That is what happens with AI coding assistant memory when the context layer gets scrambled by mixed languages.
How this breaks your WordPress workflow
You might think, “I only code in English, so why do I care?” You care because this drift causes subtle logic errors. If you are debugging a race condition in a WooCommerce checkout and your comments or variable names use non-standard terminology, you push the model into a thinner region of the AI Embedding Space. That is where what I call “vibe coding” disasters come from.
// The "Naive" Mixed Approach
// 请帮我 check 这个 checkout_order_processed hook
add_action( 'woocommerce_checkout_order_processed', function( $order_id ) {
// LLM might drift here and suggest deprecated methods
$order = wc_get_order( $order_id );
});
// The "Grounded" Engineering Approach
/**
* Refactor checkout hook for high-concurrency stability.
* Focus: Avoid race conditions in metadata updates.
*/
add_action( 'woocommerce_checkout_order_processed', 'bbioon_stable_checkout_handler', 10, 1 );A consistent engineering register keeps the model anchored in the most reliable part of the vector space. When it “drifts,” it does not just switch languages, it loses access to the better training data that holds the current best practices.
Fixing the drift
If your assistant starts acting “weird,” it is usually a vector location problem, not a “stupid” model. Here is how I handle it:
- Use strict registers: for a technical question, keep the whole prompt in the engineering register and skip the conversational filler.
- Ground with metadata: use tools like MCP (Model Context Protocol) to give the model a “map” of your codebase, which keeps the AI Embedding Space inside your project’s boundaries.
- Reset the basin: if a thread starts hallucinating or slipping into another language, start a new session. Once a vector is stuck in a bad attractor basin, it is hard to pull back out.
If this AI Embedding Space stuff is eating up your dev hours, I can take it off your plate. I’ve been wrestling with WordPress since the 4.x days.
The bottom line
The geography of AI is not what we assumed. We are working in a multidimensional space where a single filename can shift the “language” of the response. Understanding the AI Embedding Space is what separates a developer who gets real use out of AI from one who ends up stuck in a Korean-language loop while fixing a CSS bug. Keep your prompts in a clean engineering register and watch for the drift.
