As foundational models move fast in size and capability, the gap between research to production is yet to be completed. To iterate through the models and implementations faster, the toolchain matters a lot. This talk will cover the techniques and experience during this journey from the training of the vanilla foundational models, and optimization comparisons, to user-side system experience optimization.

How do you write a stress test for an LLM? This talk explores cutting-edge techniques to red-team generative AI and build validation engines that algorithmically probe models for security, ethics, and safety issues. Attendees will learn a framework to manage AI risk spanning the model lifecycle, from data collection through production.

An opinionated view of how to build production LLMs.

AI Squared is an AI platform designed for product owners, data scientists, and enterprise leaders. We empower you to accelerate both predictive and generative AI projects, measure their benefits, and drive significant revenue growth and cost reduction. The largest gap in the LLM developer stack is creating different experiences for users to leverage LLM results. Many companies have only utilized generative AI inside of chat applications. AI Squared has developed a framework that empowers our customers to harvest context and connect to additional content as well as other AI models from across the organization while integrating these insights directly into currently existing tools and applications.

The rapid adoption of Large Language Models (LLMs) has swept across various industries, inspiring companies to incorporate them into their products, regardless of the industry's domain. Unlike organizations like OpenAI, Google, or Microsoft, most entities view LLMs as powerful tools rather than standalone products. In this context, it remains paramount to uphold core product principles where customer satisfaction reigns supreme. In this talk, we will see how to pick practical use cases for using LLMs, where they are not essentially user-facing chatbots but valuable tools to build new features into existing products - with examples from the cybersecurity domain. We will also see a sample product strategy for how to plan and roadmap LLM features, along with key performance metrics to gauge success.

Time series—data ordered chronologically—constitutes the underlying fabric of systems, enterprises, and institutions. Its impact spans from measuring ocean tides to tracking the daily closing value of the Dow Jones. This type of data representation is indispensable in sectors such as finance, healthcare, meteorology, and social sciences. However, the current theoretical and practical understanding of time series hasn't yet achieved a consensus among practitioners that mirrors the widespread acclaim for generative models in other fundamental domains of the human condition, like language and perception. Our field is still divided and highly specialized. Efforts in forecasting science have fallen short of fulfilling the promises of genuinely universal pre-trained models. In this talk, we will introduce TimeGPT, the first pre-trained foundation model for time series forecasting that can produce accurate predictions across various domains and applications without additional training. A general pre-trained model constitutes a groundbreaking innovation that opens the path to a new paradigm for the forecasting practice that is more accessible and accurate, less time-consuming, and drastically reduces computational complexity. We will show how to use TimeGPT in a live demo.

In this talk, we will cover how to develop and deploy RAG-based LLM applications for production. We will cover how the major workloads (data loading and preprocessing, embedding, serving) can be scaled on a cluster, how different configurations can be evaluated and how the application can be deployed. We will also give an introduction to Anyscale Endpoints which offers a cost-effective solution for serving popular open-source models.

Conversational AI demands low latency for a seamless dialogue between humans and AI. However, engineers face the dilemma that some latency is inherently required in order to process human speech and craft a response. Some incremental wins to shave off milliseconds involve trade-offs against how the AI response could be enriched during the additional processing time. Others simply refactor out inefficiency to obtain more performant results from AI devtools. This talk presents best practices of designing streaming speech-to-text applications, as well as reasons to accept extra latency for the sake of an enhanced product experience.