GPT stands for Generative Pre-trained Transformer.
The most recent iteration of OpenAI's huge language model systems, GPT-4, is taught to anticipate the next word in a sentence by absorbing vast amounts of text from the internet and identifying patterns via trial and error.
According to the FT, the methodology has been included into a number of apps, including the Duolingo language learning programme and Morgan Stanley Wealth Management.
According to The Times, Duolingo offers customers two premium membership options that make use of the software: an AI conversation partner and a tool that explains how to remedy errors.
The software's stress testing, which outlined potential dangers related to prejudice, misinformation, privacy, and cyber security, were also disclosed by the creator.
GPT-4 is capable of expressing prejudices and different world views, but it may also produce "potentially destructive content, such as advise on preparing assaults or hate speech," making it vulnerable.
Nevertheless, whereas the publication of earlier models, such as GPT-3, provided technical information about the model architecture and data training, this is not currently included for the release of GPT-4, with competition and safety considerations reportedly being major determinants.
Describing images
With regard to picture input, the improved chatbot can respond to questions about photos and give thorough descriptions of the images that are being asked about, much like computer vision.
This component of the update appears to further demonstrate OpenAI's intention to compete with Google's search engine, which also supports image searches but does not yet have the ability to provide answers because Google's Bard chatbot service is still only available to a small group of testers.
In order to assist users who are blind or partially sighted, OpenAI has also established a partnership with Be My Eyes, a Danish start-up that matches persons with visual impairments with human volunteers.
Tags:
Technology