[Popular Science of Chinese Military Technology] Why is it said that the reliability of ChatGPT in military applications needs to be verified?

The full name of GPT is (Generative Pre-trained Transformer). The three letters correspond to three words. "Generative" means that the model has the ability to generate text, "Pre-trained" means that before the model is fine-tuned for specific tasks, it is pre-trained on a large number of text data sets, and "Transformer" refers to the basic architecture used by the model.

Pictured: ChatGPT image (Source: Baijiahao)

GPT is a text generation deep learning model trained on data available on the Internet. It is used for question answering, text summary generation, machine translation, classification, code generation, and conversational AI.

GPT has gone through four generations of development

In 2018, GPT-1 was born, which has certain generalization capabilities, including natural language reasoning, question answering and common sense reasoning, semantic similarity recognition, classification, etc. In terms of effect, GPT-1 can only be regarded as a fairly good language understanding tool, rather than a conversational AI.

In 2019, GPT-2 quietly arrived. GPT-2 did not make too many structural innovations and designs to the original network. In terms of performance, GPT-2 has a strong talent: reading summaries, chatting, continuing writing, making up stories, and even generating fake news and phishing emails.

GPT-3, which appeared in 2020, can be called a language genius. GPT-3 can complete almost most tasks of natural language processing, such as problem-oriented search, reading comprehension, semantic inference, machine translation, article generation, automatic question answering, etc.

GPT-4, which appeared this year, has opened a new era. GPT-4 is a new generation of multimodal large model that supports image and text input and correct text output. It has a strong image recognition ability, the text input limit has been increased to 25,000 words, and it supports multiple languages. The accuracy of the answer has been significantly improved, which makes the new GPT smarter. In addition, GPT-4 also opens up role-playing and personality customization capabilities.

In the military field, GPT, as a cutting-edge artificial intelligence technology, has attracted great attention from various countries and has become a new focus of the arms race among countries in the field of artificial intelligence.

In January 2023, the U.S. Defense Information Systems Agency officially included generative artificial intelligence technologies represented by GPT in the technology watch list, vigorously developed multi-source intelligence analysis tools, and collected combat data through various channels to widely strengthen decision-making support models. In the future, GPT's database will become larger, and its analytical skills will continue to improve. Its potential value cannot be ignored. GPT can generate training materials, automatically identify targets, quickly analyze intelligence, support the battlefield of public opinion, support the battlefield environment, formulate combat plans, simulate actual combat training, ensure equipment status, and improve support capabilities.

However, in practical applications, regardless of whether GPT will have human autonomy in the future, it is fundamentally different from natural persons: its creative generation mechanism is still unclear, which makes it difficult for users to verify the results or give enough trust to its output. From a military perspective, the information sources used by GPT are all visible or already public information. Whether the "old information library" can provide sufficient battlefield value needs to be judged in practical applications. In addition, in today's situation of rampant false information, whether the data in the GPT information library can be ensured to be authentic and reliable is also a problem that GPT needs to solve in military applications.

Expert profile: Weng Zongbo is a military equipment science popularization expert. He is mainly engaged in research on domestic and foreign high-tech equipment, main combat equipment of various arms, joint combat strategies and tactics, etc. He has published more than 200 articles in military newspapers and magazines such as the "People's Liberation Army Daily", "China National Defense News", and "Weapons" magazine, and has been awarded the third-class merit once.

Produced by: Popular Science of Chinese Military Technology

Author: Weng Zongbo (military science writer)

Planning: Jin He

Scientific review: Liu Xiaofeng (senior military editor)

Producer: Guangming Online Science Department