Can DeepSeek Read BaZi? What an AI Model Can Explain and What Chart Calculation Requires

A fortune reading priced at 888 to 2,888 yuan is a different proposition from trying a chatbot out of curiosity. Yet both sit inside the DeepSeek divination story unfolding in China this February. The same model name can appear on a copied prompt and a commercial offer, making the distance between experimentation and a paid service surprisingly hard to read.

The immediate story concerns that passage from curiosity to commerce. Reports published between February 10 and 13 describe a recognizable procedure, uneven results, and sellers attaching prices to generated interpretations. They document what people and vendors are doing with a new entrance to divination. They do not establish that the entrance leads to reliable knowledge of the future.

A price attached to a familiar name

Sixth Tone's February 10 report identifies an influencer using the handle Tatara who offered DeepSeek-powered readings for 888 to 2,888 yuan per session. The offer appears in coverage of merchants selling services around the model's popularity. The price is a reported offer, not an independently established measure of quality or accuracy.

That distinction gives the event its tension. A reading can be sold at a substantial price while the process behind it remains difficult for a prospective customer to assess. The amount charged tells us about the offer. It does not tell us how much individual work went into the answer, which interpretive rules were used, or whether a prediction would withstand a fair test.

The model's public identity also supplies an attractive shorthand. Naming DeepSeek identifies a technology associated with the service, but leaves the seller's own contribution unexplained. Generated text, a person's editing, and a continuing consultation are different things a customer might purchase. A price attached to the model's name can make those distinctions less visible precisely when they matter most.

Three steps into a symbolic practice

The February 13 report published by 36Kr describes users enabling DeepSeek's deep-thinking function, pasting shared prompt templates, and replacing birth details and other personal information before submitting them. It also describes prompts that specify an interpretive system, including BaZi. This is a concrete account of an entry procedure, rather than a measure of participation across an entire platform.

The copied template is doing considerable work. It names a task and tells the newcomer which details belong in it. Someone can follow the instructions without first learning how the requested system distinguishes an input, a calculation, and an interpretation. That makes the template part of the experience, not merely a neutral container for the person's question.

This helps explain what is distinctive about the episode as a product encounter. The user is not necessarily choosing a dedicated divination service at the outset. A general conversation interface can become one through instructions supplied elsewhere. The boundary between exploring a tradition and accepting another person's packaged method is crossed inside the prompt.

The answer does not always stay put

The 36Kr report describes differing responses and reports problems involving time conversion and the interpretive system requested. Its observations should remain local to the examples examined. They are not a comprehensive performance study. But they complicate the idea that a detailed response is a stable result waiting to be retrieved from a set of birth details.

There are several different problems concealed by the single word accurate. A calculation may mishandle an input. An explanation may misstate a traditional rule. A prediction may fail despite a correctly constructed chart. Each problem belongs to a different part of the process, and repairing one does not automatically repair the others.

A repeated answer is not sufficient evidence either. Consistency would make a procedure easier to inspect, but a consistently applied symbolic rule still needs separate support for claims about external events. The important distinction is between reproducing an account and verifying what that account says about a person's life. The February reports make the need for that distinction concrete without resolving it.

The company and the sellers are separate

Sixth Tone also reports that DeepSeek issued a February 6 clarification identifying its official accounts and saying it did not offer paid knowledge-sharing services. The clarification said paid groups claiming official affiliation were fraudulent. That statement concerns affiliation claims; it should not be expanded into a finding about every independent commercial use of a model.

Here the brand becomes part of the transaction in a second way. It may name the technology being used, or appear to identify who stands behind an offer. Those meanings are not interchangeable. A customer could understand the first correctly and still misunderstand the second. The company's clarification places an institutional boundary inside a marketplace built around its name.

A useful account of this episode therefore keeps the seller, the model developer, and the generated reading in separate view. Responsibility for an offer does not disappear because software helped produce it. Equally, an independent seller's promise should not silently become the developer's promise. The distinction concerns who is making the claim before it concerns whether the claim is convincing.

A story can travel farther than its evidence

The shareable response and the paid reading have something important in common: both can arrive as a finished personal narrative. That form encourages a person to read for recognition. A striking sentence is easy to remember, while an input conversion or an uncertain rule may require slower examination. This is an editorial observation about the form, not a finding about every participant's motives.

Public circulation adds another layer of selection. An account chosen for posting cannot stand for every account that was generated. A pleasing result can be meaningful to its reader without providing the unsuccessful attempts, advance definitions, and ordinary comparisons needed to assess predictive performance. A social response and a test result answer different questions.

What the entrance changes

DeepSeek's fortune-telling moment is distinctive because the entrance is portable: a template moves into a general-purpose interface, while commercial offers gather around the same recognizable name. The February reporting lets us see those mechanisms at work. It does not require a claim that every curious user believes the answer or that every seller offers the same service.