The home for technical founders, engineers & researchers.

sota

Ivan is the co-founder of Pleias, an open science LLM startup.

He is a Research Professor for Semantic Data Processing at the Center for AI (CAIRO) at the University of Applied Sciences Würzburg-Schweinfurt. His interests span computational creativity, generative models, and NLP, with a strong preference for applied, empirical, and data-driven approaches.

Over the course of an hour, Ivan will discuss the challenges of benchmarking models—where the real difficulties lie—and share his perspectives on how to align benchmarks with real-world performance better. The session will conclude with an open Q\&A, giving you the opportunity to ask him anything about this essential tool for assessing whether we are truly moving in the right direction.


Why Your Evals Are Probably Off: Benchmarks vs. Reality

Laurent is the CTO of Kyutai, an open-science research lab launched a year ago to advance and democratize artificial intelligence. He studied mathematics and computer science at École Polytechnique and earned a PhD at the intersection of cryptography and formal methods. Over the years, he has made significant open-source contributions, including two of the most widely used machine-learning libraries for Rust.

Lucky us, Laurent will host a 1-hour deep dive into how speech-to-speech AI models work, drawing on his experience with Kyutai's open-source model Moshi and its recent variant Hibiki, which performs flawless live translation from French to English while preserving voice and intonation. The session will conclude with an open Q\&A, where you’ll have the opportunity to ask him anything about this rapidly evolving field.


Speech-to-Speech AI Models Breakthroughs

Xuan-Son Nguyen, an engineer at Hugging Face, specializes in on-device large language models (LLMs) and runtime optimization, working extensively with llama.cpp, ONNX, and maintaining the GGUF / llama.cpp integration on the Hugging Face Hub.

He will host a 1-hour deep dive into the current state of on-device LLMs — exploring their advantages, performance trade-offs, and current limitations. The session will conclude with an open Q\&A during which you can ask him anything about on-device inference, model formats, and deployment challenges.

**Engineering Spicyness level: 🌶️🌶️🌶️**


The State of On-Device LLMs

We've seen the buzz and many questions surrounding MCP from Anthropic, especially following its adoption by Cursor. Wanting to respond quickly, we’ve brought together four outstanding experts to share their insights and experiences on the topic.

Louis Abraham from H Company, Philippe Mizrahi from Linkup, Thomas Payet from Meilisearch, and Raoul Ritter from Moyai.ai will try to provide a comprehensive overview and debate the pros and cons of MCP and their best guess about its future.

We hope you'll join us in large numbers to engage with them, and we’ll ensure plenty of time for an interactive Q\&A session.

**Engineering Spicyness: 🌶️**


Why Your Evals Are Probably Off: Benchmarks vs. Reality

Speech-to-Speech AI Models Breakthroughs

The State of On-Device LLMs

Anthropic MCP: Should We Care?