After the ChatGPT earthquake , OpenAI has just launched GPT4, its next-generation language model. Its capabilities surp those of the previous model, promises the San Francisco-based company. ChatGPT, published at the end of November, caused a sensation with its ability to interact naturally with humans.
By opening ChatGPT to the public, the start-up benefited from valuable feedback, which helped it design the new model. Open AI ensures in particular that the new chatbot is less likely to be manipulated by users seeking to make it spit out dangerous information, such as the recipe for making a Molotov tail for example.
“According to our internal evaluations, GPT4 responses are 40% less likely to contain factual errors than those of GPT3.5”, the model on which ChatGPT was based, ures the Californian company. The large language models – including those of OpenAI, but also of Google or of Facebook – tend to invent facts when they do not know the answer to a question. GPT4 has yet to prove itself in this regard.
Cooking recipes and conversation
Will the new chatbot appeal to a wider audience? In a press release, the company headed by Sam Altman gives examples of possible applications: finding a common niche to organize a meeting, using an image to find cooking recipe ideas (pancakes, waffles, omelets or cakes), or compose a paragraph telling the story of Cinderella with a writing constraint: each word in the story must begin with a letter of the alphabet, in order.
On this occasion, OpenAI also unveils several partnerships with companies using its technology to build their own applications. For example, the Duolingo app uses artificial intelligence to allow someone learning Spanish to conduct conversations in that language.
More information to come…