ChatGPT: what’s new in GPT-4

Expected for March 16, he finally showed up a little earlier than expected. On Tuesday March 14, OpenAI announced the release of GPT-4, its new model of artificial intelligence. And as expected, GPT-4 will make it possible to manage more than just text.

GPT-4 is multimodal

Indeed, and this is probably the most anticipated new feature, GPT-4 is multimodal. This means that it can interpret images and text (input). In contrast, GPT-4 can only respond with text (output). Concretely, the user will be able to upload an image to support comments on ChatGPT.

For the rest, OpenAI explains that GPT-4 is capable of handling more than 25,000 words of text, which enables use cases such as long-form content creation, extended conversations as well as search and analysis of documents. It can also perform creative and technical writing tasks upon users’ request, such as composing songs, writing screenplays, or learning a user’s writing style.

“In casual conversation, the distinction between GPT-3.5 and GPT-4 can be subtleexplains the company. The difference appears when the complexity of the task reaches a sufficient threshold — GPT-4 is more reliable, creative and able to handle much more nuanced instructions than GPT-3.5″. OpenAI also released a chart comparing the two versions of GPT on different types of exams.

A more secure and reliable ChatGPT?

The firm says it has worked from the start to make GPT-4 more secure. “Our mitigations significantly improved many security properties of GPT-4 compared to GPT-3.5details the company. We reduced the model’s tendency to respond to requests for unauthorized content by 82% compared to GPT-3.5, and GPT-4 responds to sensitive requests (eg, medical advice and self-harm) in accordance with our policies 29% more often “.

GPT-4 is also more reliable, according to the words of Sam Altman, CEO of OpenAI. “It is more creative than previous models, it hallucinates much less and it is less biased”, says the latter on Twitter. The firm called on 50 experts in various fields (AI, cybersecurity, biological risks, international security, etc.) to test the model in a contradictory way, and improve the results generated. GPT-4 has also been fueled by many comments (feedback) humans

Availability of GPT-4

For developers, the new language model is accessible via the OpenAI API, distributed by invitation and therefore requiring registration on a waiting list. Some privileged developers will receive these invitations this Wednesday, March 15. The public can also access GPT-4 through ChatGPT Plus, the company’s paid program, with a usage cap.