Episode Transcript
[00:00:00] Speaker A: Lara 3 brings AI into the age of experience Supported by translated Marco Trombetti, CEO of Translated, explains how a new learning by doing Training process enables Lara3 to outperform leading MT and AI systems in quality, speed and cost efficiency.
[00:00:19] Speaker B: Why is experience now key to translation? AI
[00:00:23] Speaker A: until now, AI systems followed human written rules or learned by imitating existing data such as human translations. Learning by doing is next. Inspired by how humans learn, we begin by imitating others, then pursue goals guided by feedback, and eventually develop the judgment to evaluate and improve our own work. This is the breakthrough. A model trained by imitation can at best match what it imitates. With self judgment, we can go further.
[00:00:53] Speaker B: What's new in Lara 3?
[00:00:55] Speaker A: Rather than learning only by imitating reference translations, LARA enters a final learning by doing phase. During training, it generates alternatives, receives feedback from an automatic judge grounded in professional reviewers expertise, and refines its choices over millions of iterations.
That experience is built in before the model serves users, allowing LARA to move beyond the quality ceiling of its training data and produce translations stronger than those it learned from.
[00:01:25] Speaker B: What do users gain?
[00:01:26] Speaker A: Lara 3 reduces the need for more data and additional quality estimation steps. In our evaluations, it delivered quality that leading MT and AI models could not match while also achieving the highest throughput.
The result was especially clear in enterprise localization where terminology, brand voice, style guide compliance and product context are non negotiable.
This combination of quality and speed also delivers the highest cost efficiency among all systems tested. Source translated Zlara 3 evaluations based on the WMT 2025 benchmark and real world enterprise localization content
[00:02:08] Speaker B: About Marco
[00:02:10] Speaker A: Marco Trombetti is a computer scientist, entrepreneur and investor. In 1999 he co founded Translated, which pioneered the application of artificial intelligence in the language industry. Through Translated, Trombetti led the introduction of adaptive machine translation technology and fostered a symbiotic relationship between professional translators and AI, leveraging the success of Translated and other ventures. Marco co founded PI Campus, a venture capital firm investing in AI applied startups, and PI School, an educational institution dedicated to training the next generation of AI specialists.
This article was written by Multilingual staff. Originally published in Multilingual Magazine, Issue 257, 2026.