In the informative scientific article of the Millásreggeli radio show held on June 26, 2023, Szabados Levente, senior advisor and co-founder of Neuron Solutions, explored the possible applications of ChatGPT and other language models in businesses.
The expansion of artificial intelligence is still ongoing and now also has appeared in businesses. ChatGPT is such an artificial intelligence, a language model which is capable of answering questions. Can this new tool be used in business applications, like customer service, and is it able to earn profit?
Levente thinks the answer to this question is a complex one.
A language model is a machine learning model, which has a language composing ability. Based on its context comprising of information, it can generate texts. ChatGPT is a language model with a great confidence factor, which means it will attempt to answer a question even if the subject of that question is not present in its context. In conclusion if it gets a question like this it will potentially generate an answer false in meaning. Besides this, its ability to extract information is also limited in different ways.
The latest ChatGPT model is able to browse the net. In order to test this capability, it was given a question about Neuron Solutions. It was able to find the web page of the company, but only absorbed the first page.
However even with these mistakes, ChatGPT has business value.
Several companies think in radical terms regarding the application of language models. One of these ideas is that a model like this must be all-knowing and function without the possibility of mistakes. This idea is flawed, because a language model can only use its context to do an inference, so it’s only reliable within the confines of this context.
Another limiting idea is that the setup of a language model takes hours or days of training the model, similarly to classical machine learning tendencies or it needs fine tuning to operate to company standards. This approach of fine tuning and days of training could work for large companies and corporates, but it is out of question for smaller or medium sized companies.
Fortunately, Levente reassured us that there is indeed a golden middle way. A company wishing to use a language model only has to choose its context and prompt correctly. It needs to collect relevant information it wants to share with a pretrained language model (one like ChatGPT) and make this data into a knowledge-drive that can be passed on. The model then will be able to generate answers to questions without the need of fine tuning or all-knowing aspects.
The question arises, that what types of information is safe to share with a language model. Solutions like the one presented in the previous paragraph use a third-party tool that handles information. Levente presents an answer to this problem as well.
In this case, the information handed to the third-party software must be categorised as publicly available information. The usage of such models is limited to non-vital information that are public by default. These applications are perfect for spreading information like marketing or product details.
There’s no general model for the internal communication of a company, but the open source community delivers better and better models with an accelerating pace. These might present a solution to the above-mentioned problem but are not delivered in an instantly deployable format. The company that chooses to employ such a method needs to deal with different details, like where will the model be deployed? Locally or in the cloud? What will be the expected traffic? What will be the cost of maintaining this service?
Currently, there’s no clear advantage of using either an open source or third-party model. The intended purpose will decide this choice. For publicly accessible information, the most convenient choice is ChatGPT; and private, installed language models for internal data handling.
Last, but not least there’s the concern of employees that artificial intelligence will make their jobs obsolete. Language models, like ChatGPT will only take work hours, better spent with some more useful activity, but will not replace anyone. They won’t steal anyone’s job.
In conclusion, language models are useful business tools and will improve and become more versatile as time passes.
The full dialog is available by this link.
Follow us, for we will have more themes to explore soon!