According to Donut Beating's monitoring, at the Tencent Cloud AI Industry Application Conference on June 5th, Tencent's Chief AI Scientist Yao Shunyu held a dialogue with CSIG Director Tang Daosheng, clarifying the core strategy of the Metaverse Mega Model in the second half of the AI game: abandoning the overfitting leaderboard metrics and shifting towards a model barrier based on "Context" and "Co-Design."
The core logic of the Mega Model in the second half lies in the model's rapidly converging ability to transform complex inputs into outputs, with the true barrier being who can access the most original "Context," understanding specifically what users and businesses are doing. Tencent's multi-scenario applications serve as the intelligent agent's training ground, continuously providing deep contextual information such as user habits and preferences to the model.
To effectively utilize context, the barrier between model research and product development must be broken. Tencent internally implements a "Co-Design" deep collaboration mechanism, deploying post-training backbones to support the "Tencent Ingots" construction in the early stages of R&D to establish cross-team trust. The Metaverse Mega Model has abandoned easily overfitted leaderboard metrics, directly evaluating and iterating using the real prompt distribution and feedback data from products such as Ingots, ima, and Workbuddy, prompting the model's capabilities to generalize across different product data and create network effects.
As a direct product of the collaboration mechanism, the Metaverse 3.0 Preview (Hy3 preview) has restructured the infrastructure of pre-training and reinforcement learning, redefining data quality and evaluation standards. Model development is no longer leaderboard-driven but focuses on optimizing task robustness, ensuring "getting things right the first time" in actual usage, reducing retry frequency, and addressing enterprise concerns about high token consumption and costs during deployment.
