Name: UiPath
Uploaded: 2024-12-06T12:07:39.220Z
Duration: 2 min 33 s
Description: A leading enterprise automation software company specializing in robotic process automation (RPA) and AI-driven solutions to streamline business processes.

uipath

A leading enterprise automation software company specializing in robotic process automation (RPA) and AI-driven solutions to streamline business processes.

UiPath

Hi everyone. Welcome back to our expert series on AI. In today's episode, I'm going to talk about generative AI requirements, answering the question: What do businesses need to deploy and make the most use of generative AI technology? 

Well, generative AI is really a deep learning. It's an approach to deep learning, and deep learning is a family of algorithms that require a really significant compute. These algorithms often have billions, tens of billions, hundreds of billions, maybe even trillions of parameters. They have very, very significant computer requirements, especially the best performant and the most knowledgeable of these algorithms. 

So, the hardware requirements, just to do inference on the best models can... 

Please share the AI Agent Store with friends.

Be very very uh significant. But that's only if you want to host and run the model yourself locally. Um most people who use generative AI algorithms will actually call the algorithm over the internet and use it as a service. Uh and if you're using like that, um the hardware requirements are no more than you know being able to make a kind of API call or uh using a browser. 

[Music] I think in short no. I mean the most famous application of generative AI um is the chat GPT app. Um and uh pretty much everyone can use that to write a poem, help them compose an email without any technical um knowledge or knowhow. However, um if we want to adapt and modify generative AI models or do something more sophisticated than you know compose an. 

Share the AI Agent Store with friends.

Email or help write a block of help write block of code. Then you probably will need to have uh some familiarity with these algorithms. I wouldn't necessarily call them uh kind of technical knowledge. I'd call it more user knowledge. 

Um generative AI models can hallucinate. Generative AI models might need to be kind of prompted in a specific way or given contextual knowledge to perform a specific task. I believe everyone uh can learn that very easily and it will help them get much more from this technology. 

[Music] I think we can kind of break down um AI models into two broad families: generative AI models and specialized or supervised AI models. Generative AI models are very good at creating kind of new digital artifacts. 

Please share the AI Agent Store with friends.

Writing um you know a sentence composing a line of code um synthesizing data from a very complex kind of paragraph and uh supervised models are very good at kind of repetitive tasks like um you know reading a PDF or understanding an invoice. 

So I think there's these kind of two broad uh families of algorithms. Supervised models tend to be faster, more efficient, uh more um uh consistent. 

Generative AI models tend to be more kind of powerful, more creative but can be a little bit less consistent. 

Um within those two broad families um the best approach is really to try a model that you think will perform well and then measure its accuracy. 

Uh and then roll with that. I think the first thing to remember is most generative AI models. 

Share the AI Agent Store with friends.

Were trained on open data. That's data that you can find freely available on the web. And actually, in a lot of business contexts, the information you need to solve a problem doesn't exist on the open web. 

And if that's the case, we can never get generative AI models to provide a good response in that setting unless we give them additional information. 

And it basically, there are four techniques to giving generative AI models additional information to help them solve more nuanced and specific tasks that your business may want to apply them to. 

The first is effective prompting. The second is providing relevant contextual information. The third is what's called RAG, retrieval augmented generation techniques. And then the fourth is. 

Share the AI Agent Store with friends.

Fine-tuning so if I just go through of those one by one one I think everyone who's used chat GPT has prompted a generative AI model. It's the text that you write the description of the problem that you wanted to solve. However, people often UND specify that prompt. They don't give enough specific instructions for the model to effectively solve that problem. A good prompt can often run to, you know, tens or hundreds of lines of text. 

The second approach is providing the relevant contextual information. So imagine you're creating a generative AI application to answer customer queries, maybe to a support desk. Those support desk teams might over the years have built out a manual of, uh, over, you know, a hundred, uh, responses to all of

The FAQs and have various kind of cataloges and other information. You can effectively insert some of that information into the prompt along with the task definition so the generative AI model can create better responses. So that's, you know, providing contextual information. If that contextual information becomes too large, there's too many pages, too many documents, you'll need to use an approach called retrieval augmented generation, which effectively is just augmenting that contextual information with the database. 

So that you're putting just the most kind of relevant context into that prompt when you're calling the generative AI model. The fourth and final approach and what's arguably one of the most kind of complex to.

Execute is called fine-tuning. This is where you actually change the underlying neural network which powers the generative AI model by giving it new examples of data that might be proprietary and specific to your business and allowing it to learn from that data such that it becomes kind of um better more knowledgeable in the domain in which you're operating.

Note, that there may be slight differences in text and video. Watch video for the original content.

UiPath Video