
Why Translated Won with Lara

Combining AI and Human Work 🤖 — with Marco Trombetti
Topic 4 of 9
Published Feb 20, 2025 · Clip 23:28–31:43 (8:15)
Marco attributes Lara's success first to a team that believed it was possible, then to 15 years of meticulously collected data — not just correct translations, but machine errors, reviewer corrections, error annotations, and translator-reviewer discussions. He credits the open-source community (Mistral, LLaMA) for letting them start from 70-billion-parameter models instead of pre-training from scratch, and explains how scaling laws predicted the quality achievable with more compute. He shares concrete numbers: Lara's small model cost about 3.6 million dollars in compute, and with $40-50 million they can reach human performance; Lara already makes 2.5 errors per thousand words versus a human translator's 4, putting it at the top 30 percentile, with plans to reach the top 1 percent with 10 million training hours.