任务(项目管理)
记忆
人工智能
计算机科学
集合(抽象数据类型)
互联网
自然语言处理
万维网
心理学
数学教育
程序设计语言
工程类
系统工程
作者
Dyke Ferber,Jakob Nikolas Kather
标识
DOI:10.1016/j.euo.2023.09.019
摘要
Computer-based processing of text, including summarizing, translating, and reasoning for human-written content, has long been a highly difficult task in computer science. In recent years, however, advances in artificial intelligence (AI) have yielded a new type of machine learning model known as large language models (LLMs). Among these, the OpenAI GPT-3 (generative pretrained transformer) model released in 2020 and its ChatGPT counterpart from late 2022 have attracted significant attention. For the first time, these models have achieved human-like performance in virtually any natural language processing task. These LLMs are typically trained on a broad and diverse data set that includes text scraped from the internet. This includes medical text sources, and hence contemporaneous LLMs have internal representations of medical concepts. For example, GPT-3.5 displayed remarkable knowledge and reasoning capabilities for medical problems and came close to passing the written part of the US Medical Licensing Examination (USMLE). Its successor, GPT-4, released in March 2023, passed the examination without any difficulties and was also able to provide in-depth explanations of why it gave a specific answer [ [1] Nori H, King N, McKinney SM, Carignan D, Horvitz E. Capabilities of GPT-4 on Medical Challenge Problems. arXiv preprint. https://doi.org/10.48550/arXiv.2303.13375. Google Scholar ]. This illustrates that LLMs not only memorize their training corpus but also gain an actual understanding of the underlying concepts.
科研通智能强力驱动
Strongly Powered by AbleSci AI