If we consider Google Translate (NMT) as the 3rd Gen, then AI translation based on large language models (LLM) can be called the 4th Gen.
The characteristics of the 4th Gen can be described as being able to easily learn (certain) rules for translating "terms" and "phrases" compared to the 3rd Gen.
In that sense, the 4th Gen is very similar to the 1st Gen (rule-based) machine translation.
Even in the third generation, there are functions for glossaries and transfer learning methods such as AutoML with Google Translate, which can be tuned to the company's preferences, but the side effects are often too strong and can lead to a deterioration in translation quality. The glossary function can only be used for nouns, and the transfer learning (AutoML) method was too complicated. And it was difficult to say that it was effective.
In the 4th Gen, all you need to do is register the rules for translating terms and phrases in the system, but since it is based on a large language model, the rules can be defined in the same way as when giving instructions to humans.
For example, if you want to translate "製品" to "product". You register this in the dictionary, but in the 3rd Gen, the translation "product" is fixed as a string and does not have plural conjugation, nor is it capitalized when used as a heading. In the 4th Gen, it will handle the necessary conjugation forms.
The 4th Gen allows individual translators to cultivate AI machine translation at their own level. With the rule-based machine translation of the 1st Gen, individual translators would purchase machine translation software and painstakingly register dictionaries to cultivate the results of machine translation (although it was a laborious and unproductive effort). The same can be done in the 4th Gen.
In the 4th Gen, you can also register glossaries and translation memory (TM) as past translation assets. You can also register phrase-level translation methods. And the base translation quality itself is quite good. It is possible to grow machine translation without any side effects.
The generational shift of machine translation and AI translation occurs approximately every 10 years. In 10 years, the 5th generation will emerge, and it is impossible to imagine what it will look like.
However, the next 10 years will undoubtedly be dominated by the 4th Gen of AI machine translation.