The media reported that ChatGPT passed the United States Medical Licensing Examination (USMLE) for the right to work as a doctor in the United States. ChatGPT successfully answered 88.9 percent of the test's questions and assignments out of 350 medical tasks with a passing threshold of 60 percent. As part of the scientific study, two practicing doctors evaluated ChatGPT's answers, and discrepancies in the test results were considered separately by a third expert.
The USMLE is an exam that must be passed in order to qualify to work as an attending physician in the United States and several other countries (Israel, New Zealand, Qatar). This test is considered one of the most difficult medical examinations in the world, and its passing proves the high level of knowledge and competence of the beginning doctor, for example, for pediatric students or general practitioners from other countries.
During the ChatGPT test, it was found that the neural network provided "at least one important finding" in establishing a diagnosis that was "novel, non-obvious and clinically reliable" within its correct answers.
"Achieving a passing score on this notoriously difficult expert exam without any human assistance marks a notable milestone in the development of medical AI," researchers at technology company AnsibleHealth, which conducted the exam with ChatGPT, explained to the media.
AnsibleHealth experts believe their results suggest that ChatGPT could be a valuable tool for medical education. They said that the chatbot has a partial ability to teach medicine, revealing new and non-obvious concepts that may not be available to students at first understanding.
Simultaneously with ChatGPT, the USMLE exam was given by PubMedGPT, a neural network trained solely on medical literature and subject databases. It gave only 50.8 percent of correct answers.
In early February, ChatGPT interviewed at Google for a job as a junior software engineer, Grade L3, with a salary of about $180,000 a year. In an in-house test interview, ChatGPT beat Google's LaMDA language model in key metrics, answering all the technical questions.
Developers at Open AI created ChatGPT on the Azure AI supercomputer based on the GPT-3.5 language model. Chat-bot sought to make it easy to use, correct and "human". It was trained using a large number of texts from the Internet and the system of Reinforcement Learning from Human Feedback. ChatGPT can pass the MBA exam, solve tests and create simple polymorphic viruses.
