Enhancing LLMs

Enhancing LLMs - Agenda:
17:30-18:00 Gathering and snacks
18:00-18:45 Making Text Generation With LLMs Faster - Mike Erlichson

Autoregressive LLM decoding is slow due to sequential token generation and resource bottlenecks. In this talk I discuss techniques like speculative decoding, parallel fixed-point methods, and multi-token prediction enhance efficiency by enabling parallelism and faster token generation. The talk is based on about 15 papers :)

18:45-19:30 The New Class of Meta-AI Models - Liron Izhaki-Allerhand

We'll talk about the new class of Meta-AI models: models that help us manage our AI workflows.
We'll talk about the various ways we use models to manage AI workflows like routing, selection, adjusting complexity, and creating variability.
We'll talk about new forms of models used to manage our workflow like Adaptive RAG and StructRAG.
We'll also talk about embedding models and how they're used to manage AI workflows in various ways like from semantic retrieval to creating variability.

1. Mike Erlihson is a seasoned AI professional currently leading AI development at a stealth company, leveraging his PhD in Mathematics and extensive expertise in deep learning and data science. As a prolific scientific content creator and lecturer, he has reviewed over 250 deep learning papers and hosted more than 20 recorded podcasts in the field, building a substantial following of over 5OK in LinkedIn. In addition to his professional work, Mike is committed to education and knowledge-sharing in the AI community, making complex topics accessible through his various content platforms.
2. Liron Izhaki-Allerhand, holds a Ph.D. from Tel-Aviv, was an applied scientist in Microsoft an built and led the AI in two startups. At BioCatch, Liron build the AI team and was the first to build a (working) behavioral biometric product for mobile. At Microsoft, he worked on a variety of models across different modalities. Most notably he built Microsoft emotion model that was sold worldwide as part of the cognitive services suite. At Hour One, Liron was the first employee, led the AI department, and built one of the best AI for video generation.