Holistic Evaluation of Generative AI Systems // Jineet Doshi // #280

Key takeaways

Evaluating generative AI systems is critical for building trust but remains an open challenge due to the complexity and flexibility of LLMs.
Traditional evaluation metrics may not fully apply to LLMs, necessitating new approaches tailored to their more expansive functionalities.
Integrating evaluation throughout the entire ML lifecycle is essential for ensuring model performance and reliability post-deployment.
There is an increasing need for innovative methods like using model-based evaluations (LLMs as judges) to address the complex nature of generative AI assessments.
Ongoing evaluation strategies will be integral as AI continues to evolve and tackle more complex tasks, necessitating flexibility and scalability in approaches.

Summary

Introduction to Evaluation in AI Systems

In this episode of the NLM Community Podcast, host Amit Rios welcomes Jeanne Doshi, a staff AI scientist at Intuit. They dive into the crucial topic of evaluating generative AI systems, particularly large language models (LLMs). Doshi reflects on the journey of evaluation over the past two years, highlighting the challenges and various methods for assessing AI systems as they move towards mainstream adoption.

Challenges of Evaluating Generative AI

Doshi discusses the inherent challenges of evaluating generative AI systems, noting the shift from traditional machine learning models to LLMs that produce open-ended outputs, making it difficult to establish definitive evaluation metrics. The conversation covers how evaluations in classical ML were task-specific and well-defined but are now obscured by the OpenAI's ability to perform numerous tasks, raising questions about measuring intelligence in AI.

Approaches to Evaluation

The episode categorizes evaluation techniques for generative AI into three buckets: traditional NLP methods, human-based evaluations, and model-based evaluations using LLMs as judges. Doshi emphasizes the need to adapt existing NLP methods for open-ended tasks, the challenges posed by human labeling, and the innovative approaches arising from employing other AI models for evaluating outputs.

Implementing Evaluation in ML Lifecycle

Doshi stresses that evaluation should be integrated throughout the entire machine learning lifecycle—from data selection to post-deployment. Continuous monitoring and reevaluation are vital for maintaining the quality and integrity of AI applications as they evolve. The discussion includes insights into evaluating complex systems, like RAG (retrieval-augmented generation), and balancing cost and performance in real-world applications.

Trends and the Future of AI Evaluation

In the concluding part of the episode, the speakers touch on the future of AI evaluation—especially with the rise of agentic workflows and new breadth in multimodal outputs. They discuss simulation approaches for testing and the need for adaptable evaluation strategies that will accompany advancements in AI model capabilities, highlighting the ongoing nature of these evaluation challenges and opportunities.

Show notes

Jineet Doshi is an award-winning Scientist, Machine Learning Engineer, and Leader at Intuit with over 7 years of experience. He has a proven track record of leading successful AI projects and building machine-learning models from design to production across various domains, which have impacted 100 million customers and significantly improved business metrics, leading to millions of dollars of impact.

Holistic Evaluation of Generative AI Systems // MLOps Podcast #280 with Jineet Doshi, Staff AI Scientist or AI Lead at Intuit.

// Abstract

Evaluating LLMs is essential in establishing trust before deploying them to production. Even post-deployment, evaluation is essential to ensure LLM outputs meet expectations, making it a foundational part of LLMOps. However, evaluating LLMs remains an open problem. Unlike traditional machine learning models, LLMs can perform a wide variety of tasks, such as writing poems, Q&A, summarization, etc. This leads to the question of how to evaluate a system with such broad intelligence capabilities. This talk covers the various approaches for evaluating LLMs, such as classic NLP techniques, red teaming, and newer ones like using LLMs as a judge, along with the pros and cons of each. The talk includes an evaluation of complex GenAI systems like RAG and Agents. It also covers evaluating LLMs for safety and security, and the need to have a holistic approach for evaluating these very capable models.

// Bio

Jineet Doshi is an award-winning AI Lead and Engineer with over 7 years of experience. He has a proven track record of leading successful AI projects and building machine learning models from design to production across various domains, which have impacted millions of customers and have significantly improved business metrics, leading to millions of dollars of impact. He is currently an AI Lead at Intuit, where he is one of the architects and developers of their Generative AI platform, which is serving Generative AI experiences for more than 100 million customers around the world. Jineet is also a guest lecturer at Stanford University as part of their Building LLM Applications class. He is on the Advisory Board of the University of San Francisco’s AI Program. He holds multiple patents in the field, is on the steering committee of MLOps World Conference, and has also co-chaired workshops at top AI conferences like KDD. He holds a Master's degree from Carnegie Mellon University.

// MLOps Swag/Merch

https://shop.mlops.community/

// Related Links

Website: https://www.intuit.com/

--------------- ✌️Connect With Us ✌️ -------------

Join our Slack community: https://go.mlops.community/slack

Follow us on Twitter: @mlopscommunity

Sign up for the next meetup: https://go.mlops.community/register

Catch all episodes, blogs, newsletters, and more: https://mlops.community/

Connect with Demetrios on LinkedIn: https://www.linkedin.com/in/dpbrinkm/

Connect with Jineet on LinkedIn: https://www.linkedin.com/in/jineetdoshi/

Timestamps: