Les grands modèles de langage (LLM) ont révolutionné la résolution de problèmes en langue naturelle, mais connaissez-vous leurs limites ? Les LLM ont des tailles de contexte variant considérablement, allant de quelques milliers à plusieurs millions de tokens, mais que cela implique-t-il concrètement ?

Dans cette session, nous aborderons les points suivants via exemples illustrés et démos :

Qu’est-ce que la fenêtre de contexte d’un LLM ?

Quelle est la relation entre données, tokens, performances et coûts ?

En pratique, comment peut-on pousser les LLM dans leurs limites ?

Quels sont les cas d’usage uniquement résolus grâce à un long contexte ?

Quelles sont les différences avec une approche RAG (Retrieval Augmented Generation) ?

Deep-dive into Building BAML, a programming language for LLM prompting, with live coding and production use cases across chatbots, healthcare ERP, and finance.

Deep-dive into GovEagle's approach to automating government RFP writing with LLMs, including real-world challenges and trade-offs.

Live dive into how Hadrius builds LLM-powered compliance tools for firms managing ~$2T, including architecture, data flow, and deployment considerations.

The talk explores how large language models (LLMs) have accelerated the development of linguistic features. It focuses on how to adapt feature development processes to match this rapid pace and highlights key considerations for maintaining high-quality output in a fast-evolving AI landscape.

Abstract: The talk introduces Any Compression via Iterative Pruning (ACIP), a novel approach designed to give users intuitive control over the compression-performance trade-off. ACIP uses a single gradient descent run of iterative pruning to establish a global parameter ranking, enabling immediate materialization of models of any target size. It demonstrates strong predictive performance on downstream tasks without costly fine-tuning and achieves state-of-the-art compression for open-weight LLMs, often complementing common quantization techniques.

In this session, we'll discuss the next-generation search infrastructure that gives AI agents seamless access to web information and hard-to-find intelligence. Traditional methods can't handle these new workflows, and legacy search engines - designed for human attention - aren't built for these emerging AI use cases. We will address: a)The power of web search for LLM-based applications; b) the need to avoid scraping of legacy search engines; c) How we're building a new category of "searcher" models; and d) What you can power with a web retrieval engine, including demos.

Discover how to efficiently scale AI solutions while maintaining security and cost efficiency. Learn about integrating LLMs, intelligent search, and automated tuning into your business workflows to optimise performance and impact.

Andrew Garkavyi and Lesha Levzhynskyi discuss the history and present state of shipping features across Grammarly's multiple platforms, recounting challenges and approaches from fully native to web to hybrid, and addressing overlays, assistant and chat modes in the age of LLMs.