The Illusion of Chinese Tech Giants' 'Ontology Products': What’s Really Missing?

Every time I attend a major Chinese tech company's launch, I count how many times the speaker mentions "full‑stack self‑developed", "autonomous and controllable", and "knowledge ontology". After the count, I check the GitHub dependency list, and the numbers rarely match.

In recent years, the term "ontology" has surged in China’s tech circles. Companies building knowledge graphs claim to have "ontology engines", large‑model teams boast about "ontology cognition", and chip makers talk about "computational ontology" as if attaching the word instantly upgrades a product to a breakthrough.

In reality, the "ontology products" advertised by domestic giants are almost never genuine. They conflate "using ontology‑related technology" with "owning an ontology product", much like owning a German‑made machine tool does not make you own German manufacturing.

1. Knowledge‑Graph Ontology: Lots of Graph, No Soul

Originally, an ontology (in the semantic‑web sense) is a formal axiomatic system that can answer questions such as what a "company" is, how it relates to an "organization", what logical constraints an "employment" action must satisfy, and whether contradictory data can be automatically detected.

The core of true ontology engineering is the T‑Box (concept layer) reasoning capability , not merely a massive A‑Box (data layer).

Most domestic knowledge‑graph platforms turn out to be a visual graph database plus an entity‑extraction pipeline . Users can drag‑and‑drop nodes and edges, run NER to label "Alibaba" as a "company", and perform simple path queries. However, when asked whether a schema rule like "every listed company must have a stock code" can automatically reject an entity lacking a code, the answer is usually cannot .

These platforms largely remain at the level of graphical business‑schema editing , lacking formal semantics, Description Logic reasoners, or OWL consistency checks. Logical constraints are enforced by hand‑written rule engines or ad‑hoc property‑graph tricks, because customers prioritize "quick launch" over "logical completeness" and OWL reasoning does not scale industrially.

Thus, the so‑called "ontology product" is really a data‑management tool with a pretty UI—skin without bone.

2. The Gray Area of "Full‑Stack Self‑Developed": Shell Without Core

In the era of large models, "ontology" has been broadened to mean the foundational stack: custom chips, frameworks, and models. Companies loudly proclaim "full‑stack self‑development", yet a closer look shows a mix of partial self‑development and heavy reliance on external components.

Chip Layer

Huawei Ascend and Cambricon indeed produce custom NPUs and work on CANN and MindSpore integration, but their ecosystem maturity lags behind CUDA. Other firms’ "self‑developed compute clusters" often consist of Nvidia A100/H100 GPUs, with self‑development limited to network topology and cooling solutions . Designing chips and effectively using them are separate concerns; packaging Nvidia‑based clusters as a "self‑controllable compute ontology" is a conceptual sleight‑of‑hand.

Framework Layer

PyTorch and TensorFlow remain dominant. Domestic frameworks such as MindSpore, PaddlePaddle, and OneFlow each have technical merits, yet none have achieved the ecosystem indispensability required of a true "ontology"—they lack broad third‑party library compatibility and community‑driven contributions.

Model Layer

Models like Wenxin, Tongyi, HunYuan, and Pangu are indeed self‑developed, but questions remain: does the training‑data cleaning pipeline use open‑source tools? Are the low‑level operator libraries self‑built? Is the tokenizer a custom adaptation of SentencePiece? Can the toolchain (compilation, debugging, profiling) operate independently of the Hugging Face ecosystem?

In practice, "full‑stack self‑development" often reduces to "core model self‑development + peripheral ecosystem stitching" . This is not shameful—everyone does it—but presenting it as a completely autonomous breakthrough is misleading.

3. Large‑Model Agents' "World Model": Drama Without Brain

Finally, the most mystical "ontology" claim: AI’s understanding of the world. Companies tout agents with "deep physical cognition" and "symbolic reasoning fused with neural networks", suggesting true comprehension of concepts like "cup" or "causality".

In reality, current LLMs possess an emergent statistical association , not a symbolic causal structure. For example, GPT‑4 can correctly answer "If I flip a cup, will water spill?" because the phrases "flip" and "spill" co‑occur frequently in its training data, not because it has a formal ontology of gravity, liquids, and containers.

Domestic agents rely on the three‑piece combo of RAG + function calls + prompt engineering , which merely patches the cognitive gap. They lack a genuine concept hierarchy, physical constraint engine, or symbolic world model.

Thus, what is marketed as "ontology cognition" is simply inserting knowledge‑graph nodes into prompts for the model to "guess" relevance—essentially advanced retrieval, not true ontology.

Conclusion

I am not denying the engineering excellence of Chinese firms in deployment, productization, and cost control. Their knowledge‑graph tools solve many data‑governance problems, and their models are competitive. However, I oppose the conceptual fraud of labeling ER diagrams as ontologies, cluster optimizations as autonomous chips, and prompt engineering as world models.

A genuine ontology product requires formal semantics, an irreplaceable ecosystem, and symbolic cognition—none of which domestic giants currently possess. They have the packaging, but the fire cannot be contained by paper.