Alter Everything 196: AI Model Strategy in Translation
10 snips
Oct 22, 2025 Olga Beregovaya, Vice President of AI at Smartling, shares her extensive experience in language technology, diving into the evolution from rule-based to transformer models. She highlights the advantages of purpose-built AI for translation over general models. Olga also discusses Smartling's multi-model translation stack and the operational challenges of managing various models. The conversation includes tackling biases in translation and the benefits of curated data, particularly in risk-sensitive sectors like life sciences.
AI Snips
Chapters
Transcript
Episode notes
Transformers Were The Major Breakthrough
- Transformer models were the biggest leap in language technology after statistical methods replaced rule-based systems.
- Olga Beregovaya highlights that attention-based transformers unlocked capabilities previously impossible with older approaches.
Prefer Purpose-Built Models For Translation
- Use purpose-built models for translation tasks instead of relying solely on generalized foundational models.
- Fine-tune and curate task-specific corpora to dramatically improve translation quality and predictability.
Smartling's Multi-Model, 40-Prompt Pipeline
- Smartling runs a model portfolio across Vertex AI, WatsonX, Bedrock, Azure, and OpenAI while remaining model-agnostic.
- Their R&D keeps roughly 40 prompts and an ecosystem that can swap models as newer versions prove better for specific tasks.