with its incredible ChatGPT, a chatbot capable of conducting conversations, answering questions, and composing coherent prose, poetry, and code, has forced many to rethink the potential of artificial intelligence.
OpenAI, the startup behind ChatGPT, today announced the core of a highly anticipated new version of its AI model.
The new algorithm, called GPT-4, follows GPT-3, a groundbreaking text generation model announced by OpenAI in 2020 and later adapted last year to create ChatGPT.
OpenAI said the new model scored higher on a battery of tests designed to measure the intelligence and knowledge of humans and machines. It also reduces missteps and is responsive to images and text.
However, GPT-4 suffers from the same problems that have plagued ChatGPT and has led some AI experts to cast doubt on its usefulness—including a tendency to “hallucinate” incorrect information, exhibit problematic social biases, and be less effective at giving “confrontational” information. “hint.
“While they’ve made a lot of progress, they’re clearly not to be trusted,” said Oren Etzioni, a professor emeritus at the University of Washington and founding CEO of the Allen Institute for Artificial Intelligence. “It will be a long time before any GPT will run your nuclear power plant.”
OpenAI provides several demos and benchmark data to demonstrate the capabilities of GPT-4. Not only can the new model beat the passing score on the Uniform Bar Exam used by many US states to qualify for the bar, but it scores in the top 10% of humans.
It also scored higher than the GPT-3 in subjects such as biology, art history and calculus on other exams designed to test knowledge and reasoning. It scored better than any other AI language model on a test designed by computer scientists to measure the progress of such algorithms. “In some ways, it’s more of the same,” Etzioni said. “But in a series of absolutely exciting advances, it’s more like the same.”
GPT-4 can also perform neat tricks previously seen from GPT-3 and ChatGPT, such as summarizing and suggesting edits to snippets of text. It can also do things its predecessors couldn’t, including acting as a Socratic tutor, helping guide students to the correct answers and discussing the content of photos. For example, if a photo of ingredients is provided on a kitchen counter, GPT-4 can suggest suitable recipes. If a graph is provided, it explains the conclusions that can be drawn from it.
“It does seem to gain some capabilities,” said Vincent Conitzer, a CMU professor who specializes in AI and has begun experimenting with new language models. But he said it can still make mistakes, such as coming up with nonsensical directions or providing bogus mathematical proofs.
ChatGPT has attracted public attention for its amazing ability to solve many complex problems and tasks through an easy-to-use conversational interface. A chatbot doesn’t understand the world like a human, it just answers with words that statistically predict should follow the question.
