Navigating the LLM hype: What science can tell us about where LLMs win and where they fail.

Oren Matar

Over the past two years, Large Language Models (LLMs) have taken the tech world by storm, leading many companies to rapidly adopt them in hopes of transforming their products. Yet, despite their potential, many LLM-based applications have fallen short, sometimes even causing embarrassment or public backlash. These failures often arise from overly optimistic expectations and a lack of understanding of the models' limitations. In this session, we will systematically explore the true capabilities of LLMs. Through empirical experimentation, we’ll examine key questions: Can LLMs reason, or are they just sophisticated parrots? Can they engage in abstract thinking or demonstrate planning? After several years of research, science has a lot to say about the current and projected capabilities of LLMs. With this knowledge, we will establish practical, actionable guidelines for applying LLMs effectively in products. You’ll gain insights into which use cases are best suited for LLMs, how to avoid common pitfalls, and how to align expectations with the current capabilities of these models. This session does not require a deep technical understanding of LLMs, making it ideal for anyone, from product managers to decision-makers, who is considering using these models in their applications.

Oren Matar

I am a freelance consultant specializing in AI-driven solutions, with expertise in 3D modeling and its intersections with AI, as well as advanced analysis of NLP systems. My research focuses on addressing the limitations of generative AI and developing practical frameworks to unlock its potential in both NLP and image processing.

Previously, I worked as a principal data scientist and algorithms developer, with a specialization in NLP and time series forecasting for supply chain management. Notably, I contributed to optimizing Facebook’s Prophet library by enhancing runtime efficiency through Bayesian methods. Today, I continue to bridge theory and practice, leveraging innovative approaches to help clients and collaborators solve complex challenges in AI and beyond.