ChatGPT is a current buzzword in mainstream news. Thanks to their astonishing results, Large Language Models (LLMs) and the Generative Pretrained Transformer (GPT) have crossed the borders of AI labs to quickly become mass-adopted tools. How have we got to this? What’s next? From a nice spring to a torrid summer In the editorial for the second issue of ISTI News, in 2017, Fabrizio Falchi and I wrote: “Deep Learning is not a blanket solution for all AI problems, but it is certainly a new approach to machine learning that will have a relevant impact on a wide range of applications. Deep Learning is here to stay”. 

Seven years on, we can say we were right, and maybe even too cautious, at least in view of the events of these recent months. 2017 saw the full bloom of the AI spring, which, by convention, began in 2012 with the resounding success in an image classification challenge of the Convolutional Neural Networks of AlexNet. In recent years, research on deep learning has produced a continuous progression of new network architectures, algorithms, and “tricks” (e.g., ResNet, Adam, ReLU…), which have shaped our knowledge on neural networks. In tasks such as vision and image processing, deep learning has scored a sequence of impressive results, gaining the front spot in the news for the sensational image generation capabilities of diffusion-based methods, such as Stable Diffusion.  

Read more