In recent months, the capabilities of new AI models to generate text, code, images and much more have been capturing the attention of millions of people. However, this technology has also caused concern among some experts, leading to questions about its reliability and safety. In search of answers to these questions, our Adriatic Region Head of Data, Boris Cergol, has recently given a talk delving into the potential of generative AI and its implications for business. We sat down with Boris to get some of his key thoughts – read them below and then watch the full video of his talk.
Boris, why is generative AI so successful right now?
The key developments that have propelled generative AI forward include the scaling of model sizes, the quality of training data and the available computing power. Not to be neglected is also the important concept of foundation models, which is why I give a brief introduction to them in my talk and highlight some of their advantages. For example, the ability to solve multiple different problems using the same model and unique characteristics like the intriguing concept of emergence – a sudden and unexpected jump in the models’ capability as their size increases.
The technology still has limitations, though. How can we overcome them?
Precisely, I address some of the most discussed deficiencies of text and code generation models in my talk, such as their tendency to hallucinate erroneous answers, the staleness of their data and their disappointing performance when it comes to mathematics and reasoning.
These issues can be significantly improved with the implementation of some rather unexpected techniques, however, like appending special words to the models’ instructions, granting them access to the web or programming tools, or having them pose questions to themselves. However, this isn’t simply a matter of technique; it’s also about how we perceive the models. We need to avoid the common mistake of viewing the models as sources of information and instead view them as gateways between ideas, interfaces and implementations.
How can AI influence the way we source information?
A particularly interesting application of large language models arises when they are connected to existing knowledge bases. We can achieve this using semantic search, which focuses on the contextual meaning rather than simple keywords. This enables the language models to answer user queries using factual data sources. Furthermore, this contextual meaning can also be extended across different data types, such as directly connecting images to text that best describes them. In my talk, I show different examples of emerging search engines based on these principles to demonstrate that they provide quite a distinct user experience compared to what we are accustomed to.
So, where lies the potential for businesses?
After years of development, I believe that generative AI technologies have reached the point where they can be considered commercially viable. Nevertheless, organisations should take great care when applying these technologies and strongly favour cases where errors do not result in potentially catastrophic failures.
To gain a deeper understanding of generative AI, Boris advises us to get some first-hand experience using some of the new tools and includes many practical examples in his talk. So, if you are keen to learn more about generative AI, make sure to watch his talk: