China Smart, US Dumb: LLMs Bad, MoEs Good
November 21, 2024
Okay, an “MoE” is an alternative to LLMs. An “MoE” is a mixture of experts. An LLM is a one-trick pony starting to wheeze.
Google, Apple, Amazon, GitHub, OpenAI, Facebook, and other organizations are at the top of the list when people think about AI innovations. We forget about other countries and universities experimenting with the technology. Tencent is a China-based technology conglomerate located in Shenzhen and it’s the world’s largest video game company with equity investments are considered. Tencent is also the developer of Hunyuan-Large, the world’s largest MoE.
According to Tencent, LLMs (large language models) are things of the past. LLMs served their purpose to advance AI technology, but Tencent realized that it was necessary to optimize resource consumption while simultaneously maintaining high performance. That’s when the company turned to the next evolution of LLMs or MoE, mixture of experts models.
Cornell University’s open-access science archive posted this paper on the MoE: “Hunyuan-Large: An Open-Source MoE Model With 52 Billion Activated Parameters By Tencent” and the abstract explains it is a doozy of a model:
In this paper, we introduce Hunyuan-Large, which is currently the largest open-source Transformer-based mixture of experts model, with a total of 389 billion parameters and 52 billion activation parameters, capable of handling up to 256K tokens. We conduct a thorough evaluation of Hunyuan-Large’s superior performance across various benchmarks including language understanding and generation, logical reasoning, mathematical problem-solving, coding, long-context, and aggregated tasks, where it outperforms LLama3.1-70B and exhibits comparable performance when compared to the significantly larger LLama3.1-405B model. Key practice of Hunyuan-Large include large-scale synthetic data that is orders larger than in previous literature, a mixed expert routing strategy, a key-value cache compression technique, and an expert-specific learning rate strategy. Additionally, we also investigate the scaling laws and learning rate schedule of mixture of experts models, providing valuable insights and guidance for future model development and optimization. The code and checkpoints of Hunyuan-Large are released to facilitate future innovations and applications.”
Tencent has released Hunyuan-Large as an open source project, so other AI developers can use the technology! The well-known companies will definitely be experimenting with Hunyuan-Large. Is there an ulterior motive? Sure. Money, prestige, and power are at stake in the AI global game.
Whitney Grace, November 21, 2024
Comments
Got something to say?
-
- Subscribe to Beyond Search
Feature archive
News archive
-
