The sixteenth biennial conference of the Association for Machine Translation in the Americas (AMTA) was held in Chicago between September 30 and October 2, 2024. This was a bimodal event, meaning that people could attend in person or remotely. Of the 240 people who registered, more than half attended virtually via the Cvent Internet platform, which allowed them to view the talks as they were being presented or download them later at a more convenient time. Some conference organisers establish a general theme for their conference before publishing the first call for papers. This would have been superfluous at AMTA 2024, since there was already one major topic that was on everybody’s mind: the startling performance of recent generative AI systems (such as ChatGPT) and, more broadly, their underlying large language models, not just for translation per se, but for a range of translation related activities. I use the word ‘startling’ advisedly here because, unlike dedicated MT systems (such as GoogleTranslate or DeepL), these AI-based systems have not been explicitly trained to perform translation. Yet for certain language pairs, their performance compares very favourably with the best dedicated systems. At AMTA 2024, no fewer than twenty-four of the forty-six talks explicitly mentioned generative AI or large language models (LLMs) in their title, in addition to three of the six virtual tutorials that were presented two weeks earlier. This is not to suggest that neural machine translation (NMT) has now become outdated. On the contrary, it is very much viewed as the state-of-the-art in machine translation. But people are definitely scanning the horizon, actively comparing the newest AI-based systems to the best in NMT. An AMTA conference differs from other major events in the NLP field in that the association explicitly seeks to bring together people from the different fields with a stake in translation automation: researchers, system developers, translation service providers and, of course, translators. This year, the opening keynote address was delivered by Philipp Koehn, an important figure in the development of statistical machine translation and chief organiser of the annual Conference on Machine Translation, or WMT (Workshop on Machine Translation) as it is more commonly called. WMT attracts the first two aforementioned groups and features an open competition on various MT-related tasks, targeting different language pairs and text types, with the competing systems’ output ranked by human evaluators. At AMTA, Koehn partially leaked the results of this year’s competition, revealing that systems based on LLMs fared extremely well. Koehn mentioned a number of factors that explain their impressive performance: the enormous size of these models–between ten and a thousand times larger than the models of dedicated MT systems; the fact that they are trained on full texts rather than individual sentence pairs; and their capacity to be guided by prompts that instruct them on the users’ expectations. In the second half of his talk, Koehn went on to describe efforts to make these AI-based systems, which are still overwhelmingly English-centric, more multilingual. Given the partial WMT results revealed by Koehn, and other evaluations comparing dedicated MT systems to large generative AI systems, one might expect translators and translation service providers to be flocking to the new AI-based technology. The talk presented by Kirti Vashee, entitled ‘The Evolving Path to LLM-based MT,’ addressed just this question. Paradoxically, several of the factors that Koehn pointed out to explain the impressive performance of LLMs also account for a certain hesitancy, particularly among small-to-mid size LSPs (language service providers). Training an in-house LLM, or one that is specialised for a particular client or domain, requires huge resources: …
