For many, generative AI systems like ChatGPT may seem magical. These linguistic models are first formed on a large number of data extracted from the web. They are then refined to better respond to text instructions, resulting in a fully formed final model. The number of parameters in these models is very high and the overall training process is long. Learning such a model can take up to a year on many computers.

Once trained, all the user has to do is ask a question or make a request for the model to respond. For example, he writes the appropriate computer code or writes an essay in response to a user request. The results are often creative and sometimes remarkably “human.”

It's no surprise that over 100 million people, from CEOs to students, have adopted generative models in their lives and work. However, it is important for organizations to understand how these systems work. Despite the hype, AI is neither magical nor infallible. Recognizing its limitations will help businesses deploy safe, powerful, and revolutionary AI.

Demystifying generative AI

Advances in hardware and new AI techniques such as transformers have led to the creation of major language models (LLM) of today. It's about machine learning models at large scales trained on immense quantities of textual data using huge levels of computing power.

These models are trained to Simply predict the next word, given a context of words seen For example, in the context of “time flies”, the model will try to predict the next word. Given the training data, the model can learn to give a high probability of predicting the next word as “like,” and a high probability for “and.” The model randomly selects one of these options and adds it to the context. For example, if “and” is selected, the context becomes “time flies and.”

We are now repeating this process. Ask the model to predict the distribution of the next word based on the context “time goes by and”, for which the model could give a high probability to the word “never.” In this way, The LLM generates the output word by word by sampling the distribution of predicted words.. The LLM parameters are learned to maximize the objective of predicting, as accurately as possible, the next word in the training data.

Despite the simplicity of the learning objective, LLMs demonstrate emerging understanding and reasoning skills. Arguably, it is more effective for the model to learn the language structure and relationships in our cultural and scientific information, rather than simply storing it without learning what it “means.” For example, it is more efficient for an LLM to learn the rules of arithmetic to answer a question such as “what is 27 plus 176" rather than trying to store the answer to each arithmetic question in the training data.

These LLMs have reached the point where you can say they have beaten the Turing test. They have long functional memories, with a variant of GPT-4 capable of processing 32,768 tokens—around 50 pages of text—and writing a coherent response.

As surprising as they are,Generative AI models are not sensitive or self-aware. Of course, that doesn't make them any less useful. These systems are constantly improving and have quickly proven their value in a commercial context. ChatGPT already offers numerous possibilities for automation, freeing up capabilities and improving productivity. Knowledge workers can use them to quickly summarize long documents, write emails, and even write code. Generative AI is giving people back valuable time so they can focus on the things they love and the work that matters most.

Next-gen automation

With these guarantees in place, Advances in LLMs and generative AI could enable a big leap forward in automation. Create reliable and scalable systems that change the way we work. The latest LLMs consistently outperform humans in reading comprehension. Today, leading organizations are combining the results and understanding of AI systems with the ability to take action.

The aim here is not to replace humans in their work, but to complement them. For all its sophistication, AI is struggling to automate the most complex and creative tasks without human intervention. Whether it's active learning or manual review. Thus, AI serves as an intelligent automation agent. It allows human workers to create new and more complex automations based on AI models.

About StoryShaper:

StoryShaper is an innovative start-up that supports its customers in defining their digital strategy and the development of automation solutions tailor-made.

Sources: StoryShaper, UiPath