Exploring the Future of AI & LLMs with the Research Groups of Wix, Microsoft and IBM

Wix Engineering Meetup

Monday, September 16th, 17:30
Wix Campus. TLV

About our meetups

We invite you to a one-of-its-kind meetup, where the AI research groups of Wix, Microsoft, and IBM will share their recent works around LLMs, from optimizing prompts to customizing LLMs.

Agenda

17:30 — 18:00

Gathering: Pizza & drinks

18:00 — 18:35

Gilad Barkan (Head of Data Science Guild, Wix) and Lior Sidi (AI Research, Wix)

Customizing LLMs to Enterprise Data - the Wix Journey

At the Wix Data Science group, we explore various fine-tuning techniques to adapt LLMs for Wix's unique needs.

In this talk, We'll discuss our journey, including the development of a small Wix LLM (based on LLaMa 7B) that performed better than GPT-3.5 in various Wix tasks such as: question answering, intent classification and ticket summarization. This work highlights how to create and evaluate custom LLMs for specific tasks, potentially reducing costs and complexity while maintaining accuracy.

18:35 — 19:10

Yuval Mazor (Senior Global Black Belt, Microsoft)

Constraining your LLM Generation with Guidance

Have you ever struggled with generating valid JSON from your Large Language Model (LLM)? Or perhaps you've found it challenging to produce items from a closed list? If so, you're not alone—LLMs often seem to have a mind of their own.

Enter Guidance, an open-source library by Microsoft designed to give you fine-grained control over the output of your LLM.

In this talk, we will explore how Guidance can transform your experience with LLM-based generation, making it both easy and enjoyable. We'll delve into practical scenarios demonstrating how this powerful tool can help you achieve precise and reliable outputs that match your specifications precisely. Join us to discover how Guidance can revolutionize your approach to working with large models.

19:10 — 19:45

Dr. Liat Ein-Dor (Principal Research Scientist, IBM Research) and Dr. Orith Toledo-Ronen (Research Staff Member, IBM Research)

Conversational Prompt Engineering - a User Friendly Framework for Personalized Prompt Generation

Prompts are the way we communicate with LLMs. Clear and informative prompts are crucial for guiding LLMs towards generating the desired output. However, prompt engineering can be a tedious and time-consuming process which often requires significant expertise.

In this talk we will describe Conversational Prompt Engineering (CPE), a user-friendly approach that helps users create personalized prompts for their tasks. CPE is based on a chat model that chats with users and considers their unlabeled data, to help them express their output preferences and integrate them into the prompt. In the talk, we will explain how it works, and show how clients are leveraging CPE to better accomplish their tasks.