Harnessing the Power of LLMs: Observations from ChatGPT and UiPath

Large language models example UiPath Communications Mining

Every company is an artificial intelligence (AI) company. So the old saying goes. But we’re much closer than we often realize to making this maxim a reality. For a long time, access to truly powerful AI solutions has been restricted to a lucky few. And that’s because AI is hard. Difficult to train, expensive to maintain, impossible to scale.

Historically, the skill gap to achieve game-changing AI has been too high and too expensive for most. But a new era is fast approaching. It’s driven by technological advances and new approaches to using AI in the enterprise. In many core business areas, it’s now possible to democratize AI. Making it accessible to every employee—no technical ability needed.

To understand how this is possible, you need to recognize the context enabling, and the top trends driving this leap in AI development.

Large language models: teaching machines to listen

It’s no secret that OpenAI’s ChatGPT has generated a lot of excitement recently. The chatbot’s ability to generate complete, human-like text responses opens countless new AI use cases across the enterprise. But it's taken many years to get here.

Large language models (LLM) like GPT-3 have revolutionized machines’ ability to understand human language. Trained on datasets of immense size and complexity, LLMs finally help computers understand and respond to nuanced human conversations and questions. Indeed, the SuperGLUE benchmark shows new LLMs regularly beat humans in language understanding and comprehension.

ChatGPT is just one high-profile use of this technology. Companies are starting to use LLMs as foundational models for a wide range of AI use cases. These models can ingest a vast dataset, make sense of it, and find useful patterns to share back with the business. The grouping together of related data points is called clustering. LLMs have become so accurate that they can do it without any human input.

With zero training, LLMs generate valuable insights from huge language datasets. Our LLM, which is core to UiPath Communications Mining, accurately extracts emotions and important data—like dates, order numbers, and addresses—from masses of communications like emails and tickets. But this is just the first step in the AI training process.

‘Cheating’ at AI: skipping to the last 5% of model training

Sophisticated LLMs offload many of the training requirements needed to make AI business ready. They provide the base on which businesses can build.

However, while the accuracy and understanding of these models is stunning, they’re often too generic to be useful right off the shelf. As complicated as AI development is, your typical business is even more complex. Every company is unique but interconnected. Every industry has its own specialist language. A terminology you can only learn by being immersed in that world.

The solution is to put humans back into the machine. Only human subject matter experts (SMEs) can complete the last mile of model training needed to customize AI to an exact use case. And thanks to no-code model training tools like UiPath Communications Mining, this has never been quicker or easier.

First, the LLM is fed the company’s unique and specialized data. The model finds patterns and makes predictions. Yet crucially, it shares the predictions it’s unsure of with the human SME for review and correction. The human then checks or applies the correct labels to the dataset. This helps the model to learn from its mistakes and, more importantly, learn the specifics of the business it’s embedded in.

The idea is to create a continuous feedback loop with humans at the center. The more they train the AI, the better it becomes. But the training process needs to take place within a simple, no-code environment anyone can use. In this way, labeling will become the new programming. It replaces the need for costly, complicated programming in the final sprint.

We call this process active learning and we’ve already been doing it for years. UiPath has a strong foundation in enabling employee-assisted task mining to enhance our AI models. People play a key role in helping train UiPath Document Understanding models. Flying solo, employees can also use Forms AI to easily train lightweight AI models with no coding.

More recently, our acquisition of Re:infer (now UiPath Communications Mining) brought in world-leading talent in active learning from the AI research group at University College London. With Communications Mining and Document Understanding (available via the UiPath Business Automation Platform), UiPath has a complete solution to understand and automate business communications and their attachments. It's based on powerful, user-assisted LMMs trained in a simple, no-code environment.

Achieving trustworthy AI: automated model validation

Trustworthiness has been something of a weak spot for AI. When improperly trained or validated, AI models have reinforced social biases and discriminated against customers based on age, race, and gender. With LLMs and active learning, you can now train a model to be accurate enough for business automation. But how do you account for the fairness and reliability of its decisions?

Businesses and their customers change, so AI needs to change with them. Model validation is crucial for keeping AI reliable, balanced, and fair in its predictions and decision making. But, again, model validation is hard and takes time. That’s why it’s so often neglected. It’s no surprise that 73% of AI stakeholders struggle to get executive support for responsible AI practices. But if we skimp on model validation, how can we have AI that we can trust?

UiPath has done much to enable trustworthy, ethical AI. The key has been taking the effort out of model validation through automation. Communications Mining has a feature called Model Rating, which tells the user exactly what needs to be improved. We already detect potential bias in model training and rate the accuracy of our model predictions. In our new Train functionality, UiPath even guides the user to the exact actions they need to take to solve the model’s shortcomings.

At no point does the user have to dive into the code or data of the model. There’s no time wasted on false positives or negatives. The user is guided through the complete AI training process from end to end.

A paradigm shift in AI

In the future, AI will disappear. Or rather, it'll meld into the background, rightfully leaving humans to take center stage.

Innovations in LLMs, active learning, and model validation provide a blueprint for businesses. AI development, regardless of form or application, will soon be characterized by zero code and ease of use. This will allow AI to become the ever-present canvas on which business is built. It'll be everywhere and used by everyone. Improving constantly with every interaction.

From the user perspective, machine learning will just be user experience and workflows. Everything else will be hidden. Within the next few years, we’ll arrive at a place where everyone can train and use powerful AI in their work. And when everyone in a business is using AI, it’s like taking the breaks off business success.

AI is so simple and omnipresent that the people training and using it won’t even know it’s there. Invisible AI that runs your business, empowers your employees, and delights your customers. But it still needs us humans to do its job properly!

We’re very excited to share more of our AI innovations. Register now for UiPath AI Summit to see how we’re harnessing the latest AI advances to drive exceptional customer success.

George Roth
George Roth

Technology Evangelist for Document Understanding, UiPath

Get articles from automation experts in your inbox

Get articles from automation experts in your inbox

Sign up today and we'll email you the newest articles every week.

Thank you for subscribing!

Thank you for subscribing! Each week, we'll send the best automation blog posts straight to your inbox.