ChatGPT and business application

In the last 3 days I have been asked numerous times “what do you think about ChatGPT? Any comment?”. Well, sure thing, let me share my point of view - wearing the hat of a software developer that founded a conversational AI company 10 years ago, and is currently managing multi million deals selling conversational AI projects (yes, I am talking about SentiOne , check us out).?

The most recent language model, Chat GPT was created and trained exclusively for conversational engagements. OpenAI released ChatGPT on November 30; it has since taken the tech industry and the internet by storm. And I love it - as I am super excited for our whole tech industry to be pushed forward, expand, find new use cases and most importantly - gain publicity and educate the market.?

The good

ChatGPT? model has a very well-thought-out optimization method. There is a language model trained on huge data with a large number of parameters (the context for each word has 2048 words) and a separate model for scoring answers (with many reference points). As a result, Chat GPT was optimized for answering in a way that is so impressive to so many people.?

The bad

The model, however, is a blackbox. It is a Q & A bot. You ask a question, it will respond, it generates the text, but it does not actually “understand” what it was asked about. So it's very hard to imagine any real-life applications, especially in a business scenario, such as customer service or sales. Conversational AI is currently mainly adapted in customer service teams - where the key is to go through a specific multi-step process that solves the user's problem. Those scenarios and processes are strictly defined by the company; and full control over the process and the bot messages and its responses is crucial in terms of brand image, possible complaints, etc.

Another issue with ChatGPT is that it can be misleading - providing long and nice answers but with false information. As a result, conversation with this model seems legit and real, unfortunately, once you dig deeper into its answers - they can be misleading, not accurate, factless and even spreading fake news. As the models do not really understand the meaning of their answers, they can easily provide wrong information.?

The (possibly) ugly

Training such models is extremely expensive, which is a huge barrier for adoption. This is my assumption, as OpenAI does not reveal anywhere how long it took to train such a large model, or how much hardware resources it took to train and generate responses. However, my guess is that it is A LOT. Mega pint of servers and infrastructure, which makes the whole technology so expensive, it is almost impossible to use it. The previous GPT-3 model from 2020 from OpenAI was "extremely expensive" - which was another barrier to overcome. I can only hope that going forward it will get cheaper and easier to train such large models.?

How to make it very good

What ChatGPT is missing is accurate domain knowledge. It should be retrained using datasets from specific industry/ client/ companies. If we want to see applications of such AI models, we have to understand that businesses need personalized models, trained in their language, using their language and their knowledge base with their product catalogs. Business applications of AI models need to reflect their brand voice, be a trustworthy extension of their brand.?

Businesses need personalized models, trained in their language, using their language and their knowledge base with their product catalogs. Business applications of AI models need to reflect their brand voice, be a trustworthy extension of their brand.?

I don't think anyone needs to be convinced that this is the future. Looking at SentiOne resources - datasets and team - I am convinced that no one will build a better model for the Polish language than we do. It is clear that large-scale models (like GPT) allow you to improve the accuracy of intent recognitions, and that will drive the whole industry forward.?

SentiOne is currently working on? a similar model, but one based on domain data from our key industries - banking, insurance, finance. We build our datasets from social listening - by crawling millions of online opinions and conversations online - and from chatbot implementations we have already carried out. The plan is to use those models to gain better understanding of user intentions, build bots faster (as fewer learning phrases are required), automatically test them and make bot statements more human-like. Exciting times!