Let’s Talk AI

January 7th, 2025

AI has been the buzz word for 2024, and whilst the definition of AI can be incredibly broad, most people think of GenAI when they hear the term “AI” these days. One could use GenAI to create new text, new images, new music, new videos, etc. However, the most common idea that comes to mind is something like ChatGPT, i.e. generating text. Today we’ll dive a bit into how GenAI for text generation works, what it is good for and when more traditional PredictiveAI is better suited.

A fantastic resource to dive into how GenAI for text generation (referred to as LLMs) works is Large language models, explained with a minimum of math and jargon written by Timothy Lee and Sean Trott. Essentially, LLMs work in an iterative fashion predicting the “next most likely word” (or “next most likely token”) given the text sequence submitted as the input question and the response sequence generated up until that point. Really, what those LLMs are great at is information retrieval and summarisation. Moreover, as with most things in Data Science, one of the keys to training a quality LLM is the selection and cleaning of the training data; that’s the information that will be resurfaced and summarised.

Information retrieval used to be done by physically going to libraries to source journals, books and more - a pretty time and effort intensive endeavour. It was later largely replaced by online search engines like Google and other specialised database repositories - a significant improvement in accessibility and efficiency of search. Now, LLMs like chatGPT can be tremendously useful and time-saving in providing a direct answer to our questions that summarise the information we would otherwise have had to sparse ourselves. However, it’s important to note that those LLMs are mainly resurfacing and summarising knowledge that was already available. Granted, the information surfaced could have been challenging to access and find without the help of GenAI, but it is unlikely to shed new insights that weren't already out there. 

The hype behind GenAI should not overcast the power of more traditional predictiveAI. Unlike GenAI, predictiveAI enhances our human ability to draw inferences and patterns from complex data, helping us solve highly complex problems. In the healthcare space, it is predictive AI that enabled DeepMind to solve the protein-folding problem with AlphaFold and it is predictiveAI that helps create personalised cancer vaccines - two impressive achievements. As we enter 2025, we should make an effort to be more specific when referring to AI and take time to reflect about what form of AI is best suited for the task at hand.