Robot Ed: ‘Brazilian ChatGPT’ had up to 35,000 responses – 7/9/2023 – Tech

A chatbot with artificial intelligence that conveys confidence on various subjects and responds to users as if they were a person has already been developed in Brazil. Almost 20 years ago.

Robô Ed, created by the company InBot for Petrobras, went on air in 2004 with the idea of ​​giving strength to an awareness campaign about energy consumption. But the character ended up taking another proportion.

Ed, which is actually an acronym for “Energy and Development”, worked for 12 years, and came to answer 1 million questions in a single day – the mark was usually reached over the course of a month.

Funded by the state-owned company, its development involved a team of specialists in several areas, such as programming, computer graphics, linguistics, energy and psychology. The project started with 7 employees and ended with 3 in 2016.

Before Robô Ed became successful, InBot launched in 2001 a virtual assistant for Unilever, Sete Zoom. The chatbot imitated a model, interacted with users via ICQ —a popular instant messaging program at the time— and even posed for sensual rehearsals.

This human interface (or something like that) helps hide the fact that the user is actually talking to a computer program. At Robô Ed, it was applied in a more subtle way, since it was aimed at children and youth.

The character’s charismatic design helped win over an audience that either explored the still incipient Brazilian internet or learned to do so. This playful aspect also led to the robot being used in classrooms, according to InBot, which collects dozens of mentions of Ed in academic research.

Since the launch of ChatGPT, a chatbot with generative artificial intelligence from the company OpenAI, in November 2022, the Brazilian version has won a series of honors on social networks.

“Robot Ed walked so ChatGPT could run,” joked one Twitter user. “What is ChatGPT? Is it like a Robot Ed?”, asked another.

But despite having similar goals, the operation is quite different. While Robô Ed’s database was fed by some of InBot’s seven employees, ChatGPT has automated access to “a vast amount of information collected until September 2021”, according to him. This includes texts, books, articles, websites, forums, scientific journals.

“ChatGPT is much more autonomous. Despite undergoing human oversight on right and wrong, the amount of content that was used to train it is many orders of magnitude higher. That’s many terabytes of data and a statistical understanding of language structure”, explains Rodrigo Siqueira, founder and head of technology at InBot.

While OpenAI’s chatbot dares to comment on any subject and in countless ways —whether he’s right or not—, Robot Ed had 35,000 different responses at its peak.

“It was the robot with the largest amount of content we produced. When we put it on the air, it had about 2,000 types of responses to different subjects. We kept seeing which subjects were most talked about by users and, from there, we created answers to questions he didn’t know how to answer”, explains Siqueira.

“At first it was more specific, then it became more general. We put health information, relationship tips, cooking, there was even a psychologist on the team who put content to help people”.

The chatbot even commented on national political news. Or rather, don’t comment. In 2016, the year the project was shut down, the Sheet asked Robô Ed about the impeachment of former president Dilma Rousseff. “When the subject is controversial, I avoid exposing my tin face”, she said.

The program behind Robot Ed used Natural Language Processing (NLP) techniques to understand the content of messages even if they were without accents, abbreviated and with grammatical errors. GPT-4, OpenAI’s most advanced model that powers Microsoft’s Bing search engine, too, but on a different scale.

Another similarity that Robot Ed shares with ChatGPT is its ability to attract users who want to break the rules of the game. According to the founder of InBot, there were three types of people accessing the chatbot: those who tried to offend you to see your reaction; the one who tried to make him break his own rules; and the one that spoke of serious matters.

Anyone wanting to learn how to make a homemade bomb would be disappointed, because the existing answers were already in the program. Ed wasn’t able to come up with answers on his own based on his database, and therefore he also didn’t “hallucinate”, jargon used when text AIs invent information. And it wasn’t very likely that an educational chatbot could teach that kind of thing.

InBot today continues in the same business that gave rise to Robot Ed. Among its customers are Ambev, Bradesco and the TSE (Superior Electoral Court), for which it has also developed service chatbots. The company has already started integrating the ChatGPT engine (API) into its products, and admits that the OpenAI product has rocked the virtual assistant market.

“Today there is a much greater demand for quality, because before chatbots, most of the time, were very bad. If the person wrote with a wrong handwriting, sometimes he no longer understood. Now people chatting with ChatGPT are used to have quality answers, so companies are looking for better solutions,” said Siqueira.

Despite being seen as a novelty now, in 2023, and also 20 years ago, the idea of ​​a computer that talks to a human being is old. The first application, called Eliza, was developed between 1964 and 1966 by scientist Joseph Weizenbaum, from MIT (Massachusetts Institute of Technology). In 1984, the code for a translated version of Eliza, called Carla, was published in the Brazilian computer magazine Microhobby. Since then, the area has undergone significant developments.

See the timeline of chatbots from Eliza to ChatGPT

1966 – Eliza

Developed by Joseph Weizenbaum at MIT, Eliza was one of the first rule-based conversation programs. It simulated a psychotherapist and was designed to answer simple questions based on language patterns.

1972 – Parry

Created by psychiatrist Kenneth Colby at Stanford University, Parry was a chatbot that simulated a patient with schizophrenia. It is designed to mimic the responses and behavior of a person with this condition. Some experiments put Eliza and Parry to interact with each other.

1995 – Alice

Alice (English acronym for Artificial Linguistic Computational Entity on the Internet), created by Richard Wallace, was a natural language processing chatbot available on the internet. It used code called AIML (Artificial Intelligence Markup Language) to process and generate responses.

2011 – Siri

The launch of the iPhone 4S was accompanied by Siri, a virtual assistant with artificial intelligence that is still present in Apple products. Siri uses voice natural language processing to perform tasks, answer questions, and assist with activities on company devices.

2016 – Google Assistant

Like Siri, Google’s virtual assistant integrates with the company’s various services and devices. It also uses natural language understanding and artificial intelligence to provide personalized assistance and perform tasks.

2020 – GPT-3

GPT-3 (Generative Pre-trained Transformer 3) is a language model developed by OpenAI that has been pre-trained with content from across the internet to transform a demand into a coherent and natural response. Generative because it gives the impression that something new has been created. The model is one of the largest ever created, with 175 billion parameters.

2022 – ChatGPT

ChatGPT, launched in November 2022, is the chatbot developed by OpenAI based on the GPT-3.5 model. It is capable of generating coherent responses to a wide variety of questions and commands. It became, in the following months, the parameter of possibilities of generative artificial intelligence.