header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

Yao Shunyu Interviews Tang Daosheng: Tencent's Mega Model Relies on Context and Co-Design to Compete in the Second Half of the AI Game

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.

举报 Correction/Report
RelatedRecommended
So we strove to paddle forward, sailing against the current, only to be constantly pushed back to the past.
Around 2022, Tencent's blockchain suddenly collapsed in competition with the Alibaba group and disappeared from the domestic blockchain landscape. Today, Tencent has fully retreated in the new Web3 and virtual asset fields, while Tencent employees who hav
In fact, Tencent Cloud’s cooperation with the TON Foundation is to build nodes for the TON public chain in order to achieve further development.
Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish