A developer published an 8,000-word technical tutorial last week that cuts to the core of "how AI reads"—in two steps: first, look up each token as a vector; then add a separate lesson on "position." Most business readers overlook both steps, yet they're the starting point for judging whether an AI tool is actually reliable.
What This Is
The article comes from Juejin, China's developer community, and covers the two underlying mechanisms through which large models process language.
The first is "Embedding"—mapping discrete symbols to continuous number strings. When a model receives a sentence, it first assigns each token an ID, then looks up each ID in a matrix of roughly 150,000 rows × 1,024 columns, converting each ID into a 1,024-dimensional vector. Using Qwen3-0.6B as the example, its word vector table is a floating-point matrix. The article validates this with cosine similarity (a measure of how close two vectors point in the same direction; the closer to 1, the more similar): cats and dogs score 0.42, cats and cars only 0.16. This mechanism traces back to Google's 2013 word2vec, and the formula that has circulated ever since: the vector of king minus man plus woman lands closest to queen.
The second problem is order. AI's core operation is "attention": it compares all vectors pairwise, but has zero sense of "who comes first"—a property technically known as "permutation equivariance," meaning outputs rearrange exactly as inputs do. So the model needs an extra layer of "Positional Encoding" to tell it where each token sits.
Industry View
Supporters argue: once you understand these two steps, you realize AI doesn't "truly understand language"—it's just running large-scale vector operations. That's precisely why it can be packaged into products and replicated at scale. The tech community treats this as common knowledge, but many managers still cling to the "AI is like a human brain" framing. That mismatch, they say, is one of the main reasons enterprise AI projects keep failing.
But skeptical and cautious voices exist too. One view: reducing AI to "lookup table + vector math" is oversimplification—today's large models do far more in 1,024-dimensional space than that. Others warn that this kind of popular science makes people underestimate AI, treating it as "nothing special," when in reality it can write code and serve as a research assistant—capabilities that can't be dismissed with "it doesn't understand order."
The more practical risk sits in enterprise procurement: many buyers judge by demo performance alone and never dig into the underlying mechanics. They end up with tools that "fall apart with a rephrased question." Understanding these two steps matters more than memorizing any specific technical term.
Impact on Regular People
For enterprise IT: When procuring AI tools, the underlying mechanism is lookup and vector math—which means tool reliability depends heavily on training data coverage. Any question outside that coverage produces uncontrollable answers.
For individual careers: Being able to use AI and understanding AI are two different things. Once you grasp that it "doesn't understand order," you'll notice that the same question, asked differently, can yield entirely different answers. That's the basic instinct for judging whether AI output is trustworthy.
For the consumer market: Customer service bots, writing assistants, voice assistants all run on the same underlying mechanism. When choosing a product, don't fall for "smarter AI"—check whether it delivers consistently in your specific use case.