What Large Language Models Do Well Currently
Despite the sharp criticisms of AI, it does not mean that I believe it is entirely useless. At the current stage, AI has many excellent features.
Although I am not a heavy user of large language models, I am at least a deep user. I just feel that given the amount of money spent, such as the billions of dollars invested, or the amount of electricity consumed equivalent to that of a country, the functions it implements are somewhat too expensive. It's like using a herding dog to herd sheep, which can improve efficiency, but if you have to exchange the entire world's sheep for one herding dog, it loses its original meaning.
Based on my experience, the current large language models have done better than humans in the following aspects:
Translation: This is probably the most proficient work of large language models. Since they are trained with language, they have already mastered various electronic text languages and can switch smoothly between different languages. It can be said without exaggeration that even the relatively low-level versions of large language models can be considered masters of various languages. Of course, this refers to languages with writing systems. For languages without writing systems (many ethnic minorities in China have languages but no writing systems, such as the Bai and Miao), large language models are currently powerless. To demonstrate the translation capabilities of large language models, I have also translated this article into English, and everyone can see how it translates.
Proofreading: The proofreading capability of large language models complements their translation capabilities. Since they are formed by a large number of text trainings, finding typos or grammatical errors is easy for them. This article has also undergone proofreading by a large language model, and I believe there are no typos or grammatical errors. If you ask the large language model to write some clichés or platitudes, it is also very good at it. Because such texts are common on the internet, and the clichés or platitudes generated by the large language model are sufficient to meet many needs in the workplace.
Summarization and Extraction: Large language models can also be understood as compressors of language. Therefore, if you give them a long text and ask them to extract key points, their efficiency is very high. Similarly, within the allowable length range, they can also extract the characters, events, etc., from a novel. I don't know how long the input length allowed by the current large models is. As long as it is allowed, even for a voluminous work like "Dream of the Red Chamber," they can analyze all the characters, their relationships, and the important events that occur between them.
Programming: Besides natural languages, the most numerous forms of electronic existence are programming languages. Although I don't use this much, so my experience is not deep, I have used early large language models to write simple CSS and HTML. Therefore, I believe they should do well in this area. Translation between natural languages should also exist in programming languages. For example, if you give a large language model a segment of C code and ask it to implement the same functionality in Python, it should do well.
The above functions, even without large language models, already have corresponding solutions. Whether it is translation, proofreading, summarization, or programming, large language models indeed have a larger capacity, can process or generate longer texts, and naturally, all of this comes at a cost.
Large language models do not have self-awareness. Just like herding dogs need a shepherd to manage them, large language models also need human users to "supervise" their work. This requires users to have a certain level of language knowledge. For example, I cannot judge the quality of the large language model's programming because I don't know how to program. However, from the perspective of translation and proofreading, it can be said that the emergence of large language models is enough to make translators and editors lose their jobs.