Zhipu AI Bets $5 Billion on Self-Training: The LLM Data War Enters Its Second Half

The arms race in the large language model (LLM) sector has recently shifted its focus. In the past, the competition centered on who had the most parameters or the largest GPU clusters. However, recent moves by Zhipu AI—a leading Chinese AI startup—have elevated the competition to a new dimension: who can write better "textbooks." According to reports from TechNode and the South China Morning Post, shortly after securing approximately $5 billion in funding, Zhipu previewed its upcoming GLM-6.0 model and its fully self-training approach on September 13. This marks a transition in the AI industry from "scavenging the internet" to "self-producing data."

Moving Beyond "Internet Scavenging": The Three-Step Evolution of Data Strategy

As high-quality, publicly available human text on the internet nears depletion, acquiring premium data has become a critical bottleneck. The acquisition of LLM data is evolving through three stages. The first stage is "web scraping," where AI searches for knowledge across massive amounts of web pages—akin to looking for books in a junkyard, which inevitably leads to learning incorrect or harmful information. The second stage is "human annotation," relying on massive manual labor to clean data, which is highly costly and difficult to scale. The third stage, recently unveiled by Zhipu, is "fully self-training," generating high-quality corpora through synthetic data or self-play.

In the future, when calling APIs, the "hallucinations"—where models confidently generate false information—may be significantly reduced. Because the model's "textbook" is self-authored, its logical loop is more rigorous. For everyday users, this means AI assistants will become more reliable and secure when answering complex questions.

From "Wrapping" to "Reverse Exporting": A Shift in the Ecosystem Niche

Zhipu AI has maintained a strong presence recently. On August 26, as reported by Sina Finance, Zhipu confirmed that the mysterious model "Ox Alpha," which topped online usage leaderboards, was a new generation of its GLM series, with usage once doubling that of DeepSeek (another prominent Chinese AI lab). More dramatically, on September 11, Pasquale Pillitteri reported that Multiverse Quasar 438B, ranked first in Europe for AI, was actually a compressed version of China's GLM 5.2.

Over the past few years, domestic Chinese developers were often mocked for "wrapping" (superficially rebranding foreign open-source models). Today, however, top overseas applications are relying on the weights of Chinese models. This is not merely a simple product iteration, but a substantive reversal in technological influence. Chinese LLMs are moving from "catching up at the application layer" to "exporting underlying infrastructure," beginning to take the initiative on the global stage.

Reshaping the Selection Logic for Developers

Zhipu is also making aggressive moves in the open-source arena. On August 26, it launched and open-sourced the native multimodal model GLM-5.3-Flash (320B-A18B), scoring 57 on the Artificial Analysis Intelligence Index. Shortly after, on August 29, it open-sourced GLM-5.3, positioned as the strongest in agentic coding (allowing AI to autonomously find and fix bugs like a junior programmer) and cybersecurity defense.

Consider a specific scenario: You are an independent developer selecting a foundational model for a medical consultation app. Previously, you might have defaulted to overseas closed-source models for stability. Now, faced with an open-source, native multimodal option like GLM-5.3-Flash, your decision-making path changes. You might run a test set through Ox Alpha, discovering its accuracy in processing complex Chinese medical records far exceeds expectations. Coupled with the expectation of lower hallucinations from self-training, you ultimately decide to migrate your core business to Zhipu's API.

This is the most direct impact of foundational technology upgrades on everyday users. You do not need to understand what a "Mixture of Experts architecture that only activates a subset of parameters each time" is; you only need to know that the software replies to patients more accurately, and the API call costs have dropped due to open-source competition.

Beware of the "Echo Chamber" Caused by Inbreeding

However, this multi-billion-dollar technical roadmap harbors hidden risks. It is worth noting that while fully self-training solves the problem of data depletion, it may introduce new risks. When a model only learns from data it generates itself, it resembles biological "inbreeding," making it highly susceptible to falling into an information echo chamber.

If this self-training method becomes the industry standard, then future competition will no longer be a simple stacking of computing power, but a dual barrier of "computing power + data generation algorithms." What remains to be seen is that once there is a slight deviation in the initial logic, it could be infinitely amplified during self-iteration. This could lead to a decline in the model's ability to cover long-tail knowledge of the real world, or even cause "cognitive collapse" in specific domains. How to introduce "fresh water" from the external real world while self-producing data will be a threshold that next-generation models must cross.

Key Takeaways

One-sentence summary: Backed by $5 billion in funding, Zhipu AI previews the GLM-6.0 self-training method, signaling a shift for Chinese LLMs from application catch-up to foundational tech export, though the risk of data "inbreeding" warrants caution.

Question for discussion: If you are developing an AI application, would you prioritize replacing your existing overseas APIs with domestic open-source models that feature full self-training and lower hallucinations?