Gradient Dissent: Conversations on AI How DeepL Built a Translation Powerhouse with AI with CEO Jarek Kutylowski
8 snips
Jul 8, 2025 Jarek Kutylowski, the CEO and founder of DeepL, reveals the company's journey from pioneering neural machine translation to constructing custom data centers. He discusses the challenges of AI translation, emphasizing the crucial role of human context for high-quality results. Kutylowski also highlights how small teams can successfully compete with giants like Google. The conversation dives into the evolution of speech translation, complexities in model training, and the importance of curated datasets, offering a fascinating glimpse into the future of global communication.
AI Snips
Chapters
Transcript
Episode notes
Specialized Translation Models Matter
- Specialized model architectures outperform generic ones for translation tasks.
- Translation models must balance accuracy and fluent writing, handling both source fidelity and target language creativity.
DeepL's Early GPU Data Centers
- DeepL built its own GPU data centers starting in 2017 due to lack of available GPU compute.
- Jarek Kutylowski personally experienced significant challenges acquiring and maintaining these early systems.
Avoid Per-Customer Model Retraining
- Avoid fine-tuning a separate translation model for each customer due to scalability.
- Inject context or customer-specific info dynamically during inference instead of retraining models.