AI Superstream: Large Language Models

Schedule

The time frames are only estimates and may vary according to how the class is progressing.

Shingai Manjengwa: Introduction (5 minutes) - 8:00am PT | 11:00am ET | 4:00pm UTC/GMT

• Shingai Manjengwa welcomes you to the AI Superstream.

Surbhi Rathore–Keynote: The One-Versus-Many LLM Debate—Customizing AI for Your Business (15 minutes) - 8:05am PT | 11:05am ET | 4:05pm UTC/GMT

• Should your business build proprietary large language models tailored to your needs, or leverage public solutions? Surbhi Rathore explores the engineering and data tradeoffs between customized and general AI without prescribing a one-size-fits-all solution. You'll learn frameworks for determining whether your business should invest in specialized LLMs, utilize public models, or balance both approaches based on factors like accuracy requirements, data sensitivity, and engineering resources. You’ll learn the principles to follow when crafting an optimal LLM strategy for the needs and constraints of your business and discover pragmatic approaches to customization that safely capitalize on AI's potential.
• Surbhi Rathore is CEO and cofounder of Symbl.ai, a purpose-built AI platform for unlocking conversation data at scale. She continues to champion the potential of multistructured communication data such as voice, video, and email, while democratizing developers’ access to best-in-class conversation understanding and generative AI tech. As a software developer with a more than a decade of experience building products for customer interactions in the telecom industry, she’s an advocate of making business-ready and secure AI accessible to businesses of all types. She's a strong advocate for immigrant founders and firmly believes in creating value through collaboration without borders.

Cameron Wolfe: Understanding ChatGPT from Scratch in 25 Minutes (30 minutes) - 8:20am PT | 11:20am ET | 4:20pm UTC/GMT

• Large language models have seen a recent explosion in popularity, but what makes them so effective? Cameron Wolfe shares the foundational ideas that power language models like ChatGPT, beginning with basics like transformer architecture and next-token prediction, and then explaining how modern language models are created, covering important ideas like alignment and prompt engineering from modern AI research.
• Cameron Wolfe is director of AI at Rebuy, a personalized search and recommendations platform for D2C ecommerce brands. He works with a team of engineers and researchers to investigate topics such as language model agent systems, reinforcement learning for product merchandising, personalized product ranking, and more. He earned his PhD in computer science from Rice University in Houston, Texas. In his free time, Cameron writes a newsletter called Deep (Learning) Focus that aims to make concepts from AI research more understandable and accessible.
• Break (5 minutes)

Ron Bodkin: Generative AI Agents for Analytics (30 minutes) - 8:55am PT | 11:55am ET | 4:55pm UTC/GMT

• The field of data science has seen major advancements in recent years, allowing businesses and organizations to gain new insights from their owned data. But special skills are needed to navigate the complex world of data analytics, and that limits who can access its benefits. Making data science accessible has emerged as a key goal—and now AI agents can help. By automating complex tasks like data preprocessing, feature engineering, and model selection, agents can empower a wider range of tech professionals to do complex data analysis without extensive technical skills. Large language models like GPT-4 and Llama 2 have demonstrated a remarkable ability to understand and generate high quality text and code, and this proficiency makes them uniquely suited for transforming how deep analytics is conducted on complex datasets. Ron Bodkin explores some of the questions and capabilities that can be used for LLM analytics; challenges including reliability, integration, and changing technology; and approaches to addressing these challenges.
• Ron Bodkin is cofounder and CEO of ChainML, which provides an open source platform for rapidly developing customized generative AI applications using collaborating “agents.” The platform enables the robust deployment and monitoring of generative AI models, ensuring they can be operated with confidence and accuracy. Ron has over 15 years of AI experience, specializing in data science, analytics, machine learning, and large-scale data processing. Previously, he worked in Google’s Cloud CTO office, with a focus on applied artificial intelligence; he served as VP and GM of AI at Teradata following Teradata’s acquisition of his company, Think Big Analytics; and he led AI engineering and CIO at Vector Institute.

Malte Pietsch: RAG Versus Fine-Tuning—How to Efficiently Tailor an LLM to Your Domain Data (Sponsored by deepset) (30 minutes) - 9:25am PT | 12:25pm ET | 5:25pm UTC/GMT

• Most enterprise use cases sooner or later require the integration of domain knowledge into the LLM application. Retrieval-augmented generation (RAG) and fine-tuning have emerged as two key methods. Malte Pietsch explains how to pick the right method for your use case and how to optimize performance, costs, and latency before moving to production. He also shares hands-on tips and tricks for evaluating and boosting the performance of RAG pipelines.
• Malte Pietsch is cofounder and CTO at deepset, where he builds Haystack and deepset Cloud to enable developers all over the world to use LLMs effectively in their business applications. Before founding deepset, he conducted NLP research at Carnegie Mellon University and worked for multiple startups as a data scientist. He's been crafting NLP applications for all kinds of businesses for more than eight years and is convinced that development workflow and user orientation are the key criteria for successful NLP/LLM projects.
• This session will be followed by a 30-minute Q&A in a breakout room. Stop by if you have more questions for Malte.
• Break (5 minutes)

Lucas Soares: Creating Specialized Environments for Enhanced Learning Using Generative AI (30 minutes) - 10:00am PT | 1:00pm ET | 6:00pm UTC/GMT

• Generative AI technologies such as LLMs have the potential to greatly impact education. Lucas Soares discusses and demonstrates how GenAI could help create highly specialized environments for education, the features these environments can provide to enhance learners' experience, and the necessary safeguards against the known limitations of GenAI models. Join Lucas to explore some of the potential avenues to improved learning powered by GenAI.
• Lucas Soares is a machine learning engineer who worked at K1 Digital and Biometrid, where he developed computer vision and NLP models for applications such as document verification, OCR-based applications, and recommender systems. He holds an MSc in cognitive science with a focus on AI and completed his master's thesis on the application of generative adversarial networks to predict mouse behavior. He was also a research assistant at the Champalimaud Foundation and gained experience working with Python, TensorFlow, PyTorch, and developing various ML models, including neural networks, siamese networks, convolutional neural networks, LSTMs, and genetic algorithms. Lucas periodically writes for technical publications, makes videos about machine learning on YouTube, and teaches courses on LLM technologies on O'Reilly.

Shelbee Eigenbrode: Scaling Generative AI Workloads into Production (30 minutes) - 10:30am PT | 1:30pm ET | 6:30pm UTC/GMT

• Shelbee Eigenbrode discusses some of the common challenges involved in moving from proof of concept to production with generative AI use cases. She also shares recommendations for creating operationally efficient processes across a typical generative AI project lifecycle.
• Shelbee Eigenbrode is a principal solutions architect for generative AI at Amazon Web Services and holds six AWS certifications. Her experience in technology has spanned 23 years and multiple industries, technologies, and roles. With 35 patents across various technology domains, she has a demonstrated passion for continuous innovation and using data to drive business outcomes. Shelbee’s also a cofounder of the Denver chapter of Women in Big Data.
• Break (5 minutes)

Rohit Saha: Leveraging Large Language Models to Build Enterprise AI (30 minutes) - 11:05am PT | 2:05pm ET | 7:05pm UTC/GMT

• Generative AI is poised to disrupt multiple industries as enterprises scramble to incorporate the technology into their product offerings. The primary driver of the rush to genAI has been the ever-increasing sophistication of large language models and their unprecedented capabilities. A few third-party vendors have led the development of foundational LLMs and their adoption by enterprises, but open source LLMs have made massive strides lately and can now compete or even outperform their closed source counterparts. This competition presents a unique opportunity to enterprises who are still investigating generative AI possibilities and examining how best to use LLMs to build enduring products. Rohit Saha showcases how open source LLMs fare when compared to closed source LLMs and provides an evaluation framework that enterprises can leverage to compare and contrast different LLMs.
• Rohit Saha is an applied research scientist on Georgian's R&D team, where he works with portfolio companies to accelerate their AI roadmaps, from scoping research problems to building ML models to moving them into production. His latest project entails figuring out how businesses can leverage large language models to address their needs. Rohit has been developing ML models across vision, language, and speech modalities for more than five years. He holds a master's degree in applied computing from the University of Toronto and spent two years at MIT and Brown, where he worked at the intersection of computer vision and domain adaptation.

Chenta Lee: The Dark Side of LLMs—How LLMs Can Be Manipulated to Reveal Sensitive Data (30 minutes) - 11:35am PT | 2:35pm ET | 7:35pm UTC/GMT

• Join Chenta Lee of IBM Security as he takes you deep into the heart of AI's hidden vulnerabilities, exploring the fascinating world of AI manipulation in cybersecurity. Drawing inspiration from the intricate layers of dreams within the movie Inception, he unveils the potential risks and countermeasures associated with AI's darker side. Discover how cybercriminals can influence AI models to perform unauthorized actions, and dive into real-world examples and experiments that shed light on the intricate dynamics of AI security. You’ll come away better equipped to fortify your organization against these evolving threats.
• Chenta Lee is chief architect of threat intelligence at IBM Security. He specializes in developing advanced threat detection frameworks and leads the strategic integration of threat intelligence into IBM’s cybersecurity solutions.

Shingai Manjengwa: Closing Remarks (5 minutes) - 12:05pm PT | 3:05pm ET | 8:05pm UTC/GMT

• Shingai Manjengwa closes out today’s event.