Rome – December 23, 2025
Invited to speak at ai-PULSE 2025, Europe’s leading premier AI conference, Translated CEO Marco Trombetti was asked to address the limits of today’s AI from a perspective many systems never see in production. The invitation reflects a growing recognition that translation, more than most AI applications, reveals where current architectures succeed and where they break.
Translation is one of the few domains where errors cannot be hidden behind plausible output. Every word must preserve meaning, intent, and brand voice. For this reason, translation exposes structural limits in modern AI architectures that other applications often overlook. In production environments, these limits translate directly into brand risk, loss of trust, and measurable quality costs.
In this talk, Marco focuses on the architecture behind most modern language models: autoregressive transformers. He explains why they perform well in many scenarios, yet struggle when precision, accountability, and trust are non-negotiable. He draws on decades of production experience, combining AI with continuous human review, in which machine output is challenged and corrected every day.
What you will learn from this talk
- Why translation exposes the limits of autoregressive transformers earlier than most AI applications
- How professional review effort provides a production-grounded way to measure progress toward singularity in AI
- Where current architectures break down, including tokenization ambiguity and the limits of step-by-step reasoning
- Why recent improvements are slowing, even as models grow larger and more expensive
- What research directions can address these limits, including parallel reasoning and learning by experience
Although the presentation is aimed at a technical audience, its core message is accessible to executives and localization leaders evaluating the future of AI-powered translation. The examples come from real production translation.