Generative Artificial Intelligence is a cutting-edge technology that can learn intricate behavior patterns from a database. This technology, powered by machine learning, enables advanced AI systems like ChatGPT and DALL-E to replicate content after being trained with large datasets.
AI systems can generate novel and distinctive information with each interaction, thanks to their ability to assimilate vast amounts of data. Furthermore, their technical architecture surpasses traditional learning approaches, facilitating autonomous evolution without human intervention.
Providing a substantial amount of textual data, videos, or images to be processed is essential to develop a generative AI. Once this data is fed into the system, the technology can respond to commands with potentially accurate or inaccurate answers.
Origin of generative AIs?
The advent of generative AIs can be traced back to the intersection of two major technological revolutions: Artificial Intelligence and Natural Language Processing (NLP). Thiago Pardo, a professor at the Institute of Mathematical and Computer Sciences (ICMC) at USP, explains this connection.
According to the professor, generative AI employs Transformer modeling, which involves a collection of artificial neural networks designed to be more “attentive” to the information it needs to learn. He further explains that these models use numerical representations (word embeddings) based on traditional linguistic assumptions.
Jomar Silva, NVIDIA’s developer relationship manager for Latin America, adds that generative AIs rely on neural networks developed over decades of research.
He mentions that several models, each with its unique way of processing, are available today. For example, GPT-3, an AI model designed for conversational purposes, was trained with an enormous amount of text data from the internet, approximately half a trillion words, per Silva’s explanation.
Text generators, such as ChatGPT or Bing Chat, possess a vast capacity that involves millions of parameters or “artificial neurons.” The processing time for these neurons can range from hours to months, depending on the size of the neural network and the volume of data utilized during training.
Jomar Silva also notes that newer technologies like GPT-4 receive training by integrating text and images. This approach reinforces the learning power of these AI systems, resulting in much more engaging and interesting outcomes than the previous models. The executive emphasizes that this advancement significantly improved over the earlier models people had grown accustomed to.
How is a generative AI trained?
When an AI model presents an inaccurate response, developers use a red flag to send feedback, which is incorporated to improve the system. On the other hand, if the answer is correct, a positive hit message is transmitted to reinforce the model.
This training system applies to various formats, including text, images, videos, and audio. The difference between these formats lies solely in the technical aspect of how the AI system interprets the command. For instance, if the model is designed to generate art, it must transform the text into an image to present the desired outcome.
With millions of interactions based on true or false responses, the AI system undergoes fine-tuning until it can accurately address most queries. Over time, the model’s errors and inconsistencies are corrected, bringing it closer to perfection.
The effectiveness of the training process depends on the nature of the problem the AI system intends to solve. For example, to determine whether an image depicts a cat or a dog, the image must be classified into these categories and other animal groups. Each image is labeled according to its contents, which the neural network absorbs. As Jomar Silva explains, when the AI system encounters similar images again, it immediately recognizes and makes the appropriate associations.
Application of generative AI in everyday life
While generative AI has proven effective in various industries, complex problems are yet to be solved, particularly in the medical sector. For instance, an AI could be trained to identify cancerous tumors using X-rays or MRI scans, as explained by the NVIDIA representative.
To train an AI for this purpose, the relevant pixels in the image that indicate the tumor must be identified and recorded in a dataset, which is then assembled on a training platform. However, since tumors can vary in shape, color, and texture, the process must be repeated countless times to achieve an acceptable pattern. Despite this, attaining an optimal level of accuracy may still take a considerable amount of time.
Nevertheless, ChatGPT already has a practical application in the medical field for composing medical records from data entry, showcasing its potential to aid healthcare professionals.
Creating generative AIs is an expensive and complex process that requires a significant investment of resources, which is why large companies mainly undertake it. This involves the need for specialized professionals, advanced technological infrastructure, and a considerable development period.
However, once a network has undergone training, it can go through transfer learning processes, which involve fine-tuning the system. This process is much faster and less expensive than training the network from scratch, making it accessible to companies of all sizes. Jomar Silva highlights this as good news for businesses implementing generative AI technology.
AIs must change society.
According to Professor Thiago Pardo, generative AI technology has the potential to disrupt and fundamentally change the way humans interact with machines. He questions whether ChatGPT’s level of revolution will lead to the replacement of traditional search engines and automatic translators.
The implications of generative AI could lead to a significant revolution in the content production industry. However, Pardo warns that this technology could also have harmful consequences, including ethical concerns and potentially exacerbating economic and social inequality, particularly in nations with economies based on high technology.
In the end, the responsibility for handling this technological innovation will fall on the global society. It remains to be seen whether this will be utilized to benefit the world or become a tool that is abused. At present, no expert has provided a definitive answer to this question.