Understanding the LLM landscape: An early 2025 overview

This talk gives an overview of the Large Language Model (LLM) landscape in early 2025, from the algorithmic to the commercial aspects. It is divided into five parts. The first reminds the audience of the autoregressive nature of most LLMs, details this mechanism and finishes by a quick overview of the most common architecture behind LLMs. The second part is focused on the growth in size of LLM architectures and their use as a product by providers. It finishes by showing a few aspects to consider for using and deploying open-source LLMs. The third part introduces concepts under the umbrella of test time compute. The fourth is centered around the addition of multimodality, tools and reasoning to LLMs. Finally, the fifth and last part focuses on the problem of finding rationale with a specific focus on the research we carried out in this field.