Valentine Horstmann

Product Manager Worldline Conversational Platform  

V: When I tested ChatGPT for the first time, I was blown away by the level of understanding, the structure of the response, and how it took into account the conversation history on all subjects... Its problem-solving abilities on all aspects made it feel like a real assistant!

AI-related questions are not new to us, and we regularly discuss them with our clients. Even before ChatGPT’s emergence, we had thoughts around it. For example, we integrated a Question & Answering system developed with our research and development department. This type of assistant uses generative models similar to ChatGPT but on a smaller scale. We already had quite satisfactory results in terms of rephrasing answers based on the question asked. ChatGPT was a game-changer, giving us much more advanced conversational perspectives thanks to the models being trained on a significantly larger volume of data. In a technical way, we can talk about evolution, but in a business way, it’s a real transformation for us.

Charlotte Pasqual

R&D Project Manager at Worldline Labs

C: I think we are witnessing an unprecedented acceleration. In fact, the revolution took place some years ago with the concept of “attention” and the new architecture of deep neural networks called “transformers”. The Large Language Models (LLMs) are now the latest generation of this kind of model, and we will probably see another one that is even more powerful in the coming months or years.

As Valentine mentioned, The success of ChatGPT comes first from its very intuitive and accessible interface, which proved the almost unlimited potential of Generative AI. Within a few days, everyone could become an expert in conversing with AI to get the best generated answers on a vast range of topics, and all of this for free – as user questions were being used to create a new datasets.

And if we consider the other promising Generative AI models – those which can create all sorts of contents from text prompts (images, videos, voice, and others), we can see there is a revolution, with the emergence of a large range of new tools and usages, that are still to be explored.

C: The GPT-like models bring text generation to the next level. Some complex challenges for the earlier generations of ML engines now seem easy to solve: the editorial quality, the management of the conversational context, the scope of the tasks it tackles, amongst others.

I see also a real paradigm shift with the prompt learning. Before, data scientists had to train their models on the relevant data to perform a specific task (supervised learning). Then, with the Transformers, the ML engines could build their own representation of the language based on examples (self-supervised learning). Now, you don’t try to adjust the model anymore, it remains quite generalized, but you have to learn how to write the best prompt in your context to obtain the most accurate result, and this skill set appears more easily reachable, perhaps as it is less technical.

As a result, the growing importance of crafting effective prompts highlights the value of having dedicated “prompt engineers” who can specialize in refining the information provided to AI models, ultimately improving the quality of output and ensuring the best possible performance for various applications.

C: Currently, with its rapidly changing evolution, we see new models, new actors, and new progress every week if not every day, which makes the current market very unstable, but also very promising, as everyone wants AI and LLMs in their products and services!

For banks, we foresee several kinds of impacts. First, on the customer side, with more automation and accessibility. Then, on the professional user side, a new set of assistants should appear, able to perform repetitive tasks, such as addressing frequently asked questions from customers while providing tailored and personalised responses reducing human error and moving faster.

But that’s something we were already forecasting with the earlier generation of assistants. What’s different now, is that those tools will be able to do more cognitive tasks too: as they have the potential to review data and synthetize it at a significant rate and with a better comprehensiveness, they could advise you on investment processes or business models for instance; they could analyze companies’ performance, realize some competitive intelligence, fill comparison tables, support marketing campaigns, and even reinforce KYC/AML processes or claim management… It is maybe one of the first times that machines could be in a position to partially replace intellectual work, and people with these type of jobs and skills will have to adjust to this new reality.

Generative AI is not without some other paradoxes though: we could imagine that their exceptional creativity, for instance, would help us to find new perspectives and foster innovative ideas, which could benefit many areas. And, at the same time, their need for resources and their energy consumption is a serious obstacle, as experience has shown that sometimes, simpler tools are just as good in well-defined business use cases but have much lower energy consumption. 

C: The first question we must ask ourselves is: who is developing these models and what datasets are they using? A Stanford University1 study showed that, in 2022, most of the new Generative AI engines have been released by industrial players, and that the major actors of academic research cannot follow the current timeline of publications. A significant part of these engines came from the American market as well, and we count only a few of European initiatives.

This raises concerns about the world view that is supported by these models, and about the cognitive biases that are associated with them. OpenAI explained that they used a principle of “reinforcement learning from human feedback” to reduce this issue and make their chatbot’s answers less toxic but we don’t really have information about the rules the human experts actually applied. There are some exceptions: the European engine called Bloom, for example, for which we have the details in their scientific publication on how they took care of some specific biases in the dataset they used.