Data Brew by Databricks Mixture of Memory Experts (MoME) | Data Brew | Episode 36
7 snips
Jan 10, 2025 Sharon Zhou, Co-founder and CEO of Lamini AI, specializes in optimizing AI models for better performance. She shares insights on the intriguing blend of determinism and probabilism in managing unstructured data. Zhou discusses proprietary techniques, including memory tuning to counteract model inaccuracies. The conversation also touches on lessons learned from AI deployment, specifically drawing from experiences like those of GitHub Copilot. Expect a mix of technical wisdom and personal anecdotes about entrepreneurship and societal pressures in the tech world.
AI Snips
Chapters
Transcript
Episode notes
Hallucinations and Accuracy
- Language models hallucinate because they optimize for average error across all examples, making them good at everything but perfect at nothing.
- Experts, unlike current LLMs, are perfect at specific things, like remembering names or key facts.
Agentic Workflows and LLM Enhancements
- Agentic workflows are a key use case, where AI agents, powered by LLMs, interact with external systems like databases or CRMs.
- LLMs are being tuned for expertise and memory to improve their reliability in these workflows.
Memory Tuning and MoE
- Lamini AI uses "memory tuning," a mix of high-bandwidth, high-fidelity Mixture of Experts (MoE) and Low-Rank Adaptation (LoRA), to enhance LLMs.
- This technique enables efficient internal retrieval within the model, improving accuracy and performance.