With its strangeness the ability to hold a conversation, answer questions and write coherent prose, poetry and code, chatbot ChatGPT has forced many to rethink the potential of artificial intelligence.
The startup that created ChatGPT, openAItoday announced a much-anticipated new version of the core AI model.
The new algorithm, called GPT-4following GPT-3, a groundbreaking text generation model that helps OpenAI announced in 2020was then adapted to create ChatGPT last year.
OpenAI says the new model scores higher in a series of tests designed to measure intelligence and knowledge in humans and machines. It also makes fewer errors and can respond to images as well as text.
However, GPT-4 suffers from the same problems that have plagued ChatGPT and caused some AI experts to question its usefulness—including a tendency to “hallucination” of disinformation. accurate, exhibiting problematic social biases and behaving inappropriately, or deeming characters annoying when presented with “enemy prompts.”
Oren Etzioni, professor emeritus at the University of Washington and founding CEO of AI Allen Institute. “It’s going to be a long time before you want any GPT to run your nuclear power plant.”
OpenAI is provided some demos and data from benchmark tests to demonstrate the capabilities of GPT-4. Not only was the new model able to beat the passing score on the Unified Bar Exam, which is used to evaluate barristers in many U.S. states, but it also scored in the top 10 percent of children. People.
It also scored higher than the GPT-3 on other exams designed to test knowledge and reasoning ability, in subjects including biology, art history, and calculus. And it scores better than any other AI language model in tests designed by computer scientists to gauge progress in such algorithms. “In a way, it’s more similar,” Etzioni said. “But it’s more of the same in a series of absolutely amazing advancements.”
GPT-4 can also do neat tricks seen from GPT-3 and ChatGPT before, such as summarizing and suggesting edits to paragraphs of text. It can also do things that its predecessors could not, including acting as a Socratic tutor to help guide students to find the right answers and discuss the contents of the pictures. For example, if given a photo of ingredients on the kitchen counter, GPT-4 can recommend a suitable recipe. If given a chart, it can explain the conclusions that can be drawn from it.
“Looks like it has gained some ability,” said Vincent Conitzer, a professor specializing in AI at CMU and who has begun testing the new language model. But he says it still makes mistakes, such as suggesting meaningless directions or giving false mathematical proofs.
ChatGPT has captured the public’s attention with its amazing ability to solve many complex questions and tasks through an easy-to-use conversational interface. Chatbots don’t understand the world like humans and only respond with words that it statistically predicts will follow a question.
