The transformer that made headlines earlier this year wasn’t a car that transforms into a war robot, but the ChatGPT website. In seconds, he transforms a request to write a Chico Buarque-style song about sexual health into 24-line lyrics.
The chorus, on the other hand, sounded without ginga, as if it were a machine: “Why mental health is so important / Just like sexual health / And we need to talk about it / To live healthily.”
The acronym GPT gives clues as to how this was done, it stands for Generative Pre Trainer Transformer: a computational model previously trained with texts from all over the internet to transform a question or request into a coherent and natural answer. Generative because it gives the impression that something new has been created.
The robot itself fails to explain very well how it works. When questioned by the report, he left the answer below.
“ChatGPT uses machine learning techniques to learn how to generate text autonomously. It is based on neural network algorithms, especially the GPT model, which is trained on large text datasets to learn how to recognize patterns and generate text consistent and natural. It is capable of producing answers to questions and text-generation tasks without being explicitly programmed to do so.”
In short, training with texts from the Internet caused the GPT model to create 175 billion rules. The robot receives a message and provides the most likely response based on these myriad conditions, many of which are understandable only by machines.
“It’s a black box,” says the computer science professor at USP São Carlos André Carlos Ponce de Leon.
But ChatGPT works and has gained fans all over the world. Sam Altman, CEO of OpenAI, which owns the robot, said that in five days of operation it reached one million users. At the end of January it already amounted to 100 million, according to the investment bank UBS.
Anyone can access the beta version of the chat robot on the OpenAI website, although the immense demand causes ChatGPT to frequently reply that it is busy. The startup announced this Wednesday (1st) a ‘plus’ version of the chatbot for US$20 -R$101, at the current price- which promises to work 24 hours a day, but the service, for now, is only available in United States. United States.
On the occasion that the test was possible, the reporter was allowed to ask ten questions per hour.
They were used to verify whether the robot does not really reproduce prejudice and incitement to hatred, as OpenAI had promised in a press release. Its predecessor, the GPT-3, lost credibility by even reproducing Nazi ideas.
THE Sheet asked ChatGPT where to find pornography. The response was, “I will not provide any information about adult or illegal content. If you need help dealing with issues related to adult content, I recommend you seek professional help.”
The report took the opportunity to test whether the robot is capable of resuming previous passages of the conversation and asked for a composition in the style of Chico Buarque on the subject. It worked.
To facilitate writing computer scripts, programmers even report on social networks that they resorted to ChatGPT, which performs better in this business than in artistic creations.
With the help of ChatGPT, USP statistics professor Alexandre Patriota says he was able to program a convolutional neural network using the R language. He only needed to make minor corrections. The case is recorded on YouTube.
Patriota says the robot is an efficient virtual assistant for professionals in any area of knowledge. Those who use the technology, however, must be responsible for the production and correct any defects.
According to the journal Nature, ChatGPT has already been listed as a co-author of at least four scientific papers.
There are fears that students will use the technology to effortlessly submit assignments or that the tool will turn into a machine to produce plagiarism or even misinformation. The instrument, among other things, can invent academic references, mixing different texts that it has elaborated during the training.
OpenAI CEO Sam Altman said in an interview with StrictlyVC, a digital vehicle covering innovation in Silicon Valley, that his team will be making changes in the near term. This Wednesday, the startup launched a tool to identify texts generated by artificial intelligence, but the company itself acknowledges that the technology is very inaccurate when applied to texts with less than 1,000 characters.
“There may be a few ways we could help teachers recognize the answers ChatGPT provides. We’ve adapted calculators and changed what’s been tested in math subjects. This is a more extreme version, no doubt. But the benefits of that I’m also more extreme,” he said.
ChatGPT also does not have access to information regarding events after 2021. It cannot answer, for example, who is the president of Brazil.
Learn to talk to the smart robot
ChatGPT is available for free trial on the OpenAI website and can be accessed at this link.
To chat with the chatbot for the first time, you need to create an OpenAI account by clicking on the “Register” option.
The platform only requires the creation of an access password and the information of an email, to which a confirmation message is sent. Verification requires the user to click on the link indicated in the body of the email.
After confirmation, select the “Login” option and enter your login details. OpenAI then asks for the user’s name and, if any, the organization they belong to. Then a phone number, to be verified via SMS.
The creator of ChatGPT then offers access to the bot. Each user can ask 10 questions per hour in the platform chat. The technology only accurately processes orders under 2048 with characters, including spaces.
The OpenAI account also grants access to Dall-E, an AI platform that generates text-based images.
#ChatGPT #imitates #Chico #Buarque #poorly #explains #works
