The engineering behind AI and ML.

As Large Language Models (LLMs) such as GPT and Llama become increasingly integrated into various sectors, including technology, finance, healthcare, and customer service, the need for robust, scalable, and cost-effective strategies has never been more critical. Yet, Generative AI solutions develop faster than anything we've seen before, leaving decision-makers baffled with the myriad different options to choose from on one hand, and the "need for speed" launching GenAI-based products on the other. In this talk, we'll explore the different types of LLM products, explaining which one fits who and for which use-cases. We'll also discuss some key principles to consider when planning your first LLM-based product. Through this presentation, attendees will gain a comprehensive understanding of the multifaceted approach required to successfully deploy and manage LLMs in production, ensuring they fit the product and company in the best way possible.

Generative AI models hallucinate and it’s a problem. It’s the main reason holding back consumer facing implementations. Even GPT-4 hallucinates 3% of the time and that's on general knowledge, not specific to your use case. There are over 10 reasons why models hallucinate. Solving all of them is generally regarded as over engineering, much like implementing every security measure available on the market is. In this talk Jonathan Yarkoni will take us through the different ways and reasons models hallucinate. He will explain the cost value of each solution and showcase several through demos.

LLMOps represents the convergence of DevOps practices with Large Language Model (LLM) development, marking an evolving domain that gains prominence as LLMs and Generative AI applications advance. In this lecture, Gad Benram, CTO and founder of TensorOps, will delve into the concept of LLM Gateways. These network components are important in centralizing access from LLM applications to the models themselves. The session will showcase various architectural designs and examine multiple implementations of LLM Gateways. Additionally, it will address their impact on crucial aspects such as logging, security, compliance, and the enhancement of LLM application performance.

One of the most common mistakes when training and deploying a machine learning model is making sure that the feature values align properly with time. This leads to things such as label leakage, where the value of the feature changed based on the eventual outcome. In this talk, Ron will be taking the audience on a journey through space and time, drawing on his experience from Uber, Coinbase and CloudTrucks, to demonstrate how rethinking features as a result of discrete events, and a proper data infrastructure can lead to better models and more consistent real-time performance.

In the world of GenAI and LLMs one truth remains constant: The next best model is around the corner. This pace brings a challenge for developers and companies alike - The Model FOMO. We'll be diving into this very challenge and talk about strategies for building adaptability into your LLM applications. AND... A sneak peek into the the exciting projects we're working on right now!