VK Video has hit a classic scaling wall. With a database of 500 million clips and peak loads hitting 1,800 requests per second (RPS), manual moderation became a bottleneck that threatened to stifle the product. According to Vladislav Chernyshev, Head of Search Quality at AI VK, traditional human assessors were no longer failing just due to volume, but because the logic itself became too complex. Instruction manuals evolved into multi-page tomes; staff had to check for mirrored frames, content originality, and audio-visual sync just to filter out low-quality uploads and deepfakes.
The Human Ceiling in Data Labeling
The economics of the process hit a hard physical limit: the labeling staff could produce a maximum of one million data points per month. For training modern ranking neural networks, this is a drop in the ocean. Furthermore, ML engineers require data in bursts for specific tasks, and quickly hiring or training hundreds of people for peak loads is virtually impossible in today’s labor market. VK solved this by radically pivoting from human labor to the Qwen2-VL-7B multimodal model. This shift enabled the system to evaluate relevance even for low-frequency "long-tail" queries—previously a blind spot for the company's algorithms.
Results of the VLM Shift
Moving labeling to an AI-driven pipeline eliminated bureaucratic barriers. Hypothesis testing cycles that used to take weeks were slashed to just a few days. The hybrid system, powered by a VLM assessor, prepares training sets and conducts offline search quality evaluations without lunch breaks or lapses in concentration. Effectively, the company has transformed subjective human expertise into a predictable, scalable computing process.
The VK case is a signal to the market: the era of cheap manual labeling is over. Those who fail to integrate VLM stacks into their pipelines will inevitably lose ground in search precision and the speed of algorithmic updates.
Hypothesis testing cycles reduced from weeks to days. Automation of complex checks like mirroring and audio-video sync. High-quality labeling for "long-tail" search queries. Predictable scaling without the constraints of the labor market.
The real challenge now is how quickly competitors can adapt these heavy models to their specific needs without letting infrastructure costs skyrocket.