ChatGPT: OpenAI launches more powerful artificial intelligence – 03/14/2023 – Tech

The artificial intelligence (AI) behind ChatGPT —a robot capable of generating texts on various subjects— has gained a safer and more accurate version of responses, called GPT-4, according to an announcement by startup OpenAI this Tuesday (14).

The previous model had a limit of 3,000 words per generated text — the new one delivers coherent articles with up to 25,000 words.

This feature, however, is not available in an open interface such as ChatGPT. Subscribers to the ChatGPT Plus plan, available for $20, already have access to the new technology. Other users can sign up for a waiting list to gain access to the developer version of artificial intelligence.

GPT-4 was anticipated by the technology sector as a step towards general artificial intelligence, when the machine would communicate better than humans.

According to OpenAI, GPT-4 delivers similar responses to ChatGPT, which was powered by GPT-3.5. The difference would be in details, such as less chance of making factual errors and delivering dangerous answers. ChatGPT, for example, generates biographies of personalities with gross errors and can provide information on how to synthesize dangerous substances for humans.

The new technology is 82% less likely to provide rebuttals that violate company guidelines and 45% more likely to deliver a factual response. ChatGPT, in theory, cannot reproduce hate speech or encourage violence, crimes and pornography.

These errors persist, according to the startup. The president of OpenAI, Greg Brockman, said at the traditional technology conference SXSW, in Texas, that launching artificial intelligence with some flaws is essential to advance in this field of knowledge. Google and Microsoft, which were at the forefront of developing AI products, have delayed releases to avoid risk in recent years.

OpenAI’s new robot was trained on Microsoft’s Azure AI supercomputers. Big tech invested US$ 10 billion in the startup to gain access to artificial intelligence resources.

According to OpenAI, GPT-4 training ended in August 2022. Since then, the robot has gone through a post-training process to avoid inappropriate behavior. The GPT-3.5 version installed on the Bing search engine, for example, has already flirted with a New York Times columnist, after a long conversation.

OpenAI did not disclose how many learning parameters GPT-4 has, which has access to factual information as of September 2021. The previous version, GPT-3, followed 175 billion rules to deliver the most likely answer.

New to GPT-4 is the ability to interpret images. According to the disclosure of OpenAI, it is possible to show the robot an image of a man with a hand full of helium balloons and ask what would happen if the strings that hold the balloons were cut. The AI ​​would respond that the balloons would float towards the sky.

The ability with logical challenges are the main improvement of the new robot, according to the startup. ChatGPT was able to be approved among the worst law graduates in the United States in an entrance test similar to the OAB, while GPT-4 is able to be among the top 10% of that group.

ChatGPT also had difficulties in calculating mathematical expressions, which reduced its ability in exact sciences. GPT-4 still fails in this regard, but it does substantially better in physics and chemistry tests, according to a report by OpenAI.