GPT-4, a more human, more precise AI and more immune to "hallucinations"

[tasnimsanika69]

As speculated last week , OpenAI has lifted the curtain on GPT-4, the latest version of the revolutionary artificial intelligence lurking in the bowels of ChatGPT . And it is apparently more creative and less prone to inventing data it does not know, and is also less at the mercy of prejudices than its predecessor.

OpenAI co-founder Sam Altman calls GPT-4 “our most capable and aligned model to date” and emphasizes that it is a “multimodal” model, allowing images in addition to text as prompts.

The new version of OpenAI's AI model is also capable of handling massive text prompts and can remember and act on more than 20,000 words entered at once (the equivalent of a short novel).

GPT-4 is now available for users of ChatGPT Plus, the paid version of OpenAI's popular chatbot .

OpenAI has also teamed up with commercial partners to offer third-party services enhanced with GPT-4 . Language israel number screening learning app Duolingo, for example, will launch a new subscription that will provide English-speaking users with AI-generated conversations in French or Spanish. Duolingo subscribers will also be able to rely on GPT-4 to have AI explain language errors they make when speaking.

Online payment platform Stripe will also use GPT-4 to answer questions from the company's corporate users and detect potential scammers on its forums.

GPT-4 is a visual and linguistic model simultaneously
During the presentation of the new GPT-4 on Tuesday, Greg Brockman, president and co-founder of OpenAI, explained the "superpowers" of AI when it comes to feeding on images . However, this function is not yet available to the general public and is currently being tested exclusively by the company Be My Eyes .

This feature allows the machine to analyze and respond to user-supplied images and to answer questions and perform tasks based on these images. “GPT-4 is not only a linguistic model, it is also a visual model,” Brockman emphasizes. “It flexibly accepts inputs that mix text and images,” the OpenAI executive adds.

GPT-4 also provides solutions to many of the criticisms that its predecessor had in the past , namely that it had no qualms about inventing data when it did not know the exact answer to a question. This problem, known as "hallucination", was compounded by other difficulties such as offensive or abusive responses when the model was confronted with erroneous prompts.

OpenAI has also improved GPT-4 to provide users with more sensitive and thoughtful responses to questions directly related to medical advice, self-harm, or content not permitted on the platform.

Despite the improvements, GPT-4 will not be immune to “hallucinations ,” OpenAI warns. However, the new AI scores 40% higher in tests aimed at measuring its propensity to engage in “hallucinations,” the company notes.

AI has also, to a certain extent, emancipated itself from the clichés that once hampered this system . And if the old GPT insisted that “you can’t teach an old dog new tricks,” the new GPT-4 will not cling to this assertion (not necessarily true) and will assure that it is indeed possible to teach new tricks to an elderly dog.