大模型关键技术与未来发展方向——从ChatGPT谈起

Key Technologies and Future Development Directions of Large Language Models:Insights From ChatGPT

大规模预训练模型,也被称为“基座模型”或“大模型”,目前被认为是通用人工智能技术的核心引擎,已经成为了全球科技竞争焦点。本文归纳总结了以聊天生成预训练转换器(Chat Generative Pre-trained Transformer, ChatGPT)为代表的生成式大模型技术研究现状和发展趋势,从大模型基座、大模型人类偏好对齐、大模型推理与评价、多模态大模型、大模型安全可控五个方面讨论了当前大模型研究的现状和挑战,并结合我国人工智能研究特点,简要分析了大模型未来的重点发展方向。

Large pre-trained models,also known as"foundation models"or"large models",are the core technical support for generative artificial intelligence models,and have become a key focus in global technological competition.This paper summarizes the current status and development trends of large model technology represented by ChatGPT.The main challenges faced by large model research are explored from five aspects:the foundation of large models,alignment of large models with human preferences,large model inference and evaluation,multimodal large models,and safety and control of large models.Based on the characteristics of China's AI field,the potential key development directions for the large model field in the future are also analyzed.