Beyond the buzzword – What are the pros and cons of ChatGPT?

Although the language model used by ChatGPT is the biggest language model created so far, the technology behind it is not revolutionary. The neural network architecture using so-called "transformers" ("GPT" stands for Generative Pre-trained Transformers) was invented by Google researchers in 2017, and since then, it has been used successfully in hundreds of Large Language Models. However, ChatGPT was a great marketing success. And it has made many people think more about applying AI in various fields, including business.

AI is a buzzword for some companies to improve their image as modern and innovative firms to attract investors and improve their valuations. Or they use AI as a marketing tool, hoping that putting "powered by AI" in their product description will boost sales.

Of course, if it is wisely and carefully used, artificial intelligence can bring tangible business benefits not only by reducing human effort and supporting people in performing their tasks. Or even completely replacing people in some standard, repetitive activities. Many companies have started using AI to improve and automate their processes, for example, in marketing, sales, accounting, HR, or customer service.  

Developing a state-of-the-art Large Language Model requires terabytes of text data and a cluster of thousands of graphics processing units (GPUs) working for several months. Training requires as much energy as a small town uses weekly. The cost of training can run into millions of dollars. However, many pre-trained Large Language Models are available as open-source software that can be downloaded from the Internet and, after some fine-tuning, can be deployed for many uses, including commercial purposes. Moreover, in many applications, the AI solution can be much simpler.

Implementing an AI system usually starts with a proof of concept that verifies both the technical feasibility and business benefits. It also helps to decide the scope of the final deployment. A successful AI system does not need to be perfect or completely replace humans in a specific area. It is pretty common for people and AI to work side by side. For example, in accounting, the AI ​​system analyzes the texts of an incoming invoice and assigns the correct account, cost center or tax code based on learned invoice posting patterns in the past. This approach can work for most recurring invoices, but in exceptional cases where the system cannot predict where to post the invoice based on historical data, the invoice is routed to accountants.

What is dangerous about AI are the unrealistic expectations, especially at the current stage of AI development. An AI system may work relatively well in most cases. However, in some cases, its operation may become unpredictable. Therefore, its recommendations should not be the sole determiner of important decisions.

An AI system deployed to production should be closely monitored and when its performance diminishes, it should be re-trained or decommissioned. And this needs to be planned well in advance. Artificial Intelligence is just a tool. A complex tool that can be very helpful if correctly used. But it can be harmful if applied in a wrong way. or not supervised closely enough.