OpenAI unveiled the latest version of chatbot on Tuesday, boasting its GPT-4 was “safer” and less likely to stray from facts, but still faced limitations.
“GPT-4 is 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses” than earlier versions, OpenAI said, while warning it was not yet fully reliable.
The new chatbot will cost users US$20, accessed through the paid version of ChatGPT, offering OpenAI’s “most advanced system yet,” which it said scored in the ninetieth percentile on the US bar exam.
Unlike previous models, GPT-4 can process both images and text, a tool OpenAI said could be used to help blind and partially sighted people.
“Despite its capabilities, GPT-4 has similar limitations to earlier GPT models: it is not fully reliable, has a limited context window, and does not learn from experience,” OpenAI added.
“Care should be taken when using the outputs of GPT-4, particularly in contexts where reliability is important.”
Its free-to-use predecessor, GPT 3.5, became the fastest-ever application to reach 100mln users, passing the milestone just two months after its November release.
This prompted a host of big tech firms, including Microsoft Corporation (NASDAQ:MSFT) – which invested US$10bn in OpenAI in January - and Google owner Alphabet Inc (NASDAQ:GOOG) to launch their own chatbots which will feature in products such as Bing and Google Cloud.
OpenAI also said it had embedded GPT-4 in a range of software already, including language-teaching app Duolingo and Morgan Stanley (NYSE:MS) Wealth Management, to aid employees retrieve information.
The UK's intelligence agency warned of the dangers of chatbots on Tuesday though, telling users to "take great care" with the information they submit, given operators will "most certainly" access it.