28/02/2023

'Chat GPT, talk about IMPA', asks master's student Daniel Perazzo.

ChatGPT é uma Open IA (Inteligência Artificial) | Foto: Pexels

Months before packing his bags to move from Recife to Rio de Janeiro, computer engineer Daniel Perazzo started a casual conversation about IMPA, the institute where he will pursue his master's degree in 2023. His interlocutor wasn't a fellow graduate student or researcher, but rather ChatGPT, an artificial intelligence (AI) tool. The institute's mathematicians are already testing the AI's capabilities, which has been recognized for simulating natural language when generating written content.

According to Perazzo's account , the tool correctly predicted most of the information about IMPA and Visgraf (IMPA's Computer Graphics Laboratory), which the student will also be a part of this year.

"Until, surprisingly, he got many things right. The model accesses the internet or trains on a database that contained information about Visgraf," said the master's student.

Launched in November of last year, ChatGPT has become one of the most talked-about topics in the technology field. The tool has been touted as a possible replacement for Google and has received a billion-dollar investment from Microsoft.

The acronym stands for Generative Pretrained Transformer, which is code capable of being "trained" to generate text content from a database through reinforcement learning. There's a lot of technology behind OpenAI that's being explored by mathematicians.

One of the curious individuals is Luiz Velho, lead researcher at Visgraf, who emphasized ChatGPT's communication capabilities.

“The paradigm of how humans interact with machines is changing. More intelligent interfaces are getting closer and closer to how humans communicate with other humans. And now we've reached the point where the interface has changed. The great challenge for society as a whole is understanding how to deal with this,” he explained.

The ease of communication is explained by the tool's learning method, which uses deep neural networks. Machine learning works from computational models inspired by the functioning of the human brain, which then adjust after a success or failure.

“ChatGPT was trained to converse. So there were already these great language models – Siri, Alexa, and several others – that knew how to answer a question, but couldn't converse. So they [ChatGPT] started to shift to reinforcement learning, using conversations between humans as a reference. And the technique allows the tool to improve itself as it is used,” said Luiz Velho.

Advances in machine learning are also applied to image simulation programs. On a dual screen, Bernardo Alevato, also a collaborator at the Visgraf laboratory, tests another artificial intelligence program, Stable Diffusion , which creates images from an example entered by the user. He uploads a photo of his own face and asks for a pair of Ray-Ban sunglasses. The new image appears on the screen in less than 30 seconds with a classic model from the brand.

“All of this is still under development, but these are tools that seek the most natural interface and whose performance can be evaluated. As you have fewer processes to reach the result, you can better evaluate that result. We can see the best way to 'ask' or 'ask' the tool,” said Alevato.