Paras Qureshi has over 20 years of experience as a product leader in the financial industry. Currently, she is responsible for Moody’s Investors Service data and utility solutions, rating and analytic capabilities in service of structured transactions and broad AI enablement across the global credit rating agency.
They are not.
Artificial intelligence broadly leverages machines to replicate human behavior and be able to break down steps and act out different functions that can be replaced with machines.
Machine learning is able to then actually look into the relationships that are hidden in the data and build out models of those real systems. Then you're able to actually then start building out algorithms and patterns that go beyond your basic artificial intelligence, deep learning and you're starting to get into neural networks.
Neural networks mimic how the human brain works and figure out that if I'm trying to solve a particular problem, I'm going to need information from these various datasets. It pulls them together in a way that makes sense.
Generative AI goes a step further than all of that and it's actually able to then create new content that has not existed before. It's completely brand new. But it's based on those patterns that information that has been trained upon.
Generative AI itself creates new content. But it is based on foundational models that are those large machine models. And it's got a large amount of unlabeled unstructured data that it is able to leverage to be able to come up with various responses. Large language models are literally relying on natural language processing, being able to take all of that text and correspond that to the outputs. Foundational model because they are able to take broader unstructured data, unlabeled data and work with that, chat GPT essentially, it's just your application sitting on top of all of that. So when we talk about how Gen AI can be used, where we're seeing it it is things like text, summarization, text generation, and Q&A.
Now we are moving forward into text recognition, being able to extract information, learning personalization, especially in the e-commerce world, and being able to create personalized experiences based on human behavior and interactions in the past.
AI is not plug and play. AI isn't; generative AI definitely isn't. Because AI itself is not plug and play, you don't want to take it off the shelf and simply implement it. There are guardrails that are needed. There is a lot of work that's needed to make sure that the outputs that you're getting is what you actually desire. Because when it comes to bias, the data set that you're training it on t carries whatever biases it carries. Iif you're looking at financial lending decisions that carry certain biases historically. If you're now, building a solution, using AI, to reach financial lending decisions even faster, that's great. It's not going to weed out the bias, though that has occurred historically. Same with your recruiting practices; it's not going to weed out the bias if you've leaned towards a particular demographic and your hiring practices. When a solution is being trained, or a model is being trained on the historical information, it's going to amplify that bias. So it's very important to recognize that the LLM is not solving for those problems. It's it's spitting out the solution or the answer that makes the most sense based on the information that you've given it thus far.
