Anand Kannappan (@anandnk24, CEO @PatronusAI) talks about evaluating AI models for hallucinations, managing data quality, automating the process, and optimizing models.
SHOW: 927
SHOW TRANSCRIPT: The Cloudcast #927 Transcript
SHOW VIDEO: https://youtube.com/@TheCloudcastNET
CLOUD NEWS OF THE WEEK: http://bit.ly/cloudcast-cnotw
NEW TO CLOUD? CHECK OUT OUR OTHER PODCAST: "CLOUDCAST BASICS"
SPONSORS:
SHOW NOTES:
Topic 1 - Welcome to the show, Anand. Give everyone a quick introduction.
Topic 2 - Our topic today is Preventing AI Model Hallucinations. Before we dig into that, I wanted to ask about your time as Lead Data Scientist at Meta. What was it like to be early into that organization, and what did you take away from your time there?
Topic 3 - Ok, let’s dig into model evaluations and hallucinations. Let’s start at the beginning. How do model hallucinations come about?
Topic 4 - When evaluating models for hallucinations, how does a developer or a data scientist know fact from fiction? Due to its size, complexity, and number of parameters, it’s not feasible to simply fact-check and manually verify inputs to outputs. How is this process evaluated and automated with some level of confidence? Additionally, numerous benchmarks are available. What are your thoughts on the usefulness of the benchmarks?
Topic 5 - How does the concept of data quality play into this? How would we know when a model was given insufficient or improper data vs. a hallucination
Topic 6 - We often hear about how frontier models are running out of training data, and increasingly, synthetic data is being used. Does this impact hallucinations in any way?
Topic 7 - The last item I wanted to ask you about, Partonus AI, also pertains to model optimization. Can you explain that process?
FEEDBACK?