One of Arthur C. Clarke’s famous ‘three laws’ is: “Any sufficiently advanced technology is indistinguishable from magic”. As an early science-fiction writer and futurist, Clarke envisioned we’d have interstellar travel, crewed flights to Saturn, and technologies like suspended animation by 1999. Obviously, many of these predictions didn’t come true. But we’ve gotten very close to his vision of artificial intelligence (AI).
To many people, generative AI systems like ChatGPT and GPT-4 can seem like magic. These language models are trained initially on a very large set of data scraped from the web. They are then fine-tuned to better respond to text-based instructions, resulting in the final fully trained model. The number of parameters of these models is very large and the overall training process is lengthy. It may take up to a year of running on many computers to train such a model.
Once trained, the user simply asks a question or makes a request and the model responds. For example, writing the appropriate computer code or writing an essay in response to a user request. Its outputs are often creative and at times remarkably ‘human’.
It’s no surprise that over 100 million people, from CEOs to students, have embraced generative models in their lives and work. However, it’s important that organizations understand how these systems work. Despite the hype, AI isn’t magic, nor is it infallible. Recognizing its limitations will help businesses deploy AI that's safe, powerful, and revolutionary.
Advances in hardware and new AI techniques such as transformers have led to the creation of today’s large language models (LLMs). These are large-scale machine learning models trained on immense quantities of text data (essentially a significant proportion of all text data on the internet) using enormous levels of computing power.
These models are trained to simply predict the next word, given a context of previously seen words (or “tokens,” which are smaller parts of words). For example, given the context “time flies”, the model will try to predict the next word. Given the training data, the model might learn to give a high probability to predict the next word as “like,” and high probability for “and.” The model randomly selects one of these options and adds this to the context. For example, if “and” is selected, the context now becomes “time flies and”.
We now repeat this process. Asking the model to predict the distribution of the next word based on the context “time flies and,” for which the model might give a high probability to the word “never.” In this way, the LLM generates the output word by word by sampling from the distribution of predicted words. The parameters of the LLM are learned to maximize the objective to, as accurately as possible, predict the next word in the training data.
Despite the simplicity of the training objective, LLMs demonstrate emergent understanding and reasoning skills. Arguably, it’s more efficient for the model to learn the structure of language and the relations in our cultural and scientific information, rather than simply storing them without learning what they “mean.” For example, it’s more efficient for an LLM to learn the rules of arithmetic to answer a question like “what’s 27 plus 176” rather than trying to store the answer to every arithmetic question in the training data.
These LLMs have reached the point that they have arguably conquered the Turing test. They have long functional memories, with one variant of GPT-4 able to process 32,768 tokens—about 50 pages of text— and write a coherent response.
As amazing as they are, generative AI models aren’t sentient or self-conscious. Of course, this doesn’t make them any less useful. These systems are constantly improving and have rapidly proven their value in a business context. ChatGPT already provides many automation opportunities, freeing up capacity and enhancing productivity. Knowledge workers can use them to quickly summarize long documents, draft emails, and even write code. Generative AI is giving precious time back for people to focus on the things they enjoy and the work that matters most.
While generative AI models are impressive, they don’t work correctly 100% of the time. This can pose real challenges for a business. For example, Google Bard wiped $100 billion off the value of its parent company after making a simple factual error.
Just like us humans, models make mistakes. Businesses need to make sure models operate reliably, fairly, legally, and efficiently. That won’t happen unless businesses take a systematic approach to using generative models. We don't have all the answers yet, but there are ways to help businesses adopt AI safely and responsibly.
When generative AI is deployed in enterprise workflows, these models shouldn't be left unattended. Generative models are prone to factual errors and ‘hallucinations’, where the system seems to misinterpret user instructions or answer an unrelated question. Confused and even offensive outputs sometimes do occur as well.
For this reason, having a human in the loop is useful. Before a customer is exposed to a model’s output, someone in the business should have reviewed it, removing any offensive or erroneous content. This information can also be fed back to retrain the model and minimize future inappropriate generations. Ultimately, your human employees remain the best safeguard against the mistakes of AI models. The more important or high-risk the decision being made by AI is, the greater the need for human oversight.
An LLM can do some impressive things out of the box. However, it won’t necessarily understand the details of your business, your constraints, or business know-how that’s uniquely valuable to the company. The traditional approach to this problem has been supervised learning. Pretrained machine learning models are refined and fine-tuned by teams of subject matter experts (SMEs). However, any amount of model customization requires retraining the model on your personal datasets. This is expensive for the business and time-consuming for its SMEs, pulling them away from more valuable work.
The active learning approach is a potential solution. In active learning, SMEs are only asked for input when the model is unsure how to interpret a data point. This way, SMEs need only help interpret the most important data points, while the system continues to learn and re-train in the background. In this way, a business can create custom LLMs tailored to their exact needs. Furthermore, they can train them faster than traditional approaches, without sacrificing performance or accuracy.
AI is often adopted piecemeal, with businesses gradually introducing point solutions as needed. However, businesses need to consider how individual AI components will interact with other systems in the enterprise. Do they have strong integrations with workflow, or enterprise resource planning (ERP) systems? A lack of prebuilt connectors may mean AI outputs can’t be properly actioned, limiting their wider benefit to the enterprise.
More importantly, are these AI systems secure? Privacy and the prevention of data leakage are key, especially when models handle personal data. Systems like ChatGPT rely on user-shared data to improve. Any information you provide could become part of the knowledge base and might be shared with other users.
A business may benefit from taking a platformized approach instead. Many of the latest AI capabilities are combined within a number of enterprise solutions, including the UiPath Business Automation Platform. Preexisting connectors to other systems come as standard. What’s more, cloud-based platforms are kept up to date with the latest security policies and known vulnerabilities.
With these safeguards in place, advances in LLMs and generative AI could enable a great leap in automation. Creating reliable and scalable systems that change the way we work. The latest LLMs consistently outperform humans in reading comprehension. Now leading organizations are combining the output and understanding of AI systems with the ability to take action.
Let’s take the example of a customer service function. A customer email containing an attachment comes in. The customer request is processed by the type of custom LLM which orchestrates how to respond to the email, which steps and software processes need to be taken, and with what information. For example, an intelligent document processing solution takes care of the attachment, but the LLM realizes it needs to do a database lookup based on information contained in the attachment. The LLM then prepares an email response to the customer, which is checked by a human. Finally, the LLM generates a call to update the CRM.
The purpose here isn't to replace humans at work, but to augment them. For all its sophistication, AI struggles to automate the most complex and creative work without human input. Whether that’s through active learning or manual review. In this way, AI serves as an intelligent automation agent. Empowering human workers to create new and more complex automations using AI models as the foundation.
UiPath devotes considerable research to making enterprise AI safe and effective. For more insight, watch our on-demand session recordings from UiPath AI Summit.
Sign up today and we'll email you the newest articles every week.
Thank you for subscribing! Each week, we'll send the best automation blog posts straight to your inbox.