Open AI presents GPT-4

Open AI presents GPT-4

San Francisco, 3/17/2023

OpenAI introduces GPT-4, its latest and most advanced large language model that accepts both text and images. GPT-4 is now publicly available through OpenAI’s ChatGPT Plus paid subscription. The latest GPT has also already powered Microsoft’s Bing search engine AI, the company announced.

While GPT-4 accepts both text and images as input, its output remains text only.

For example, a user might upload a photo or image, and GPT-4 responds with a text description or caption for what it „sees“.

OpenAI has also started a waitlist for a GPT 4 API for developers.

GPT-4 early adopters include Duolingo, Morgan Standley, Khan Academy and Stripe.

GPT 4 improvements:

The new GPT is stronger than GPT-3.5 which supported ChatGPT.

GPT-4 supports up to 25,000 text words, an increase of 1,000.

The model is said to be „safer and better aligned“ than previous models. In internal testing, it was 40% more likely to generate factual information compared to GPT 3.5.

According to OpenAI, GPT-4 performed well on various tests and benchmarks. It was administered in the 93rd percentile of an SAT reading test and in the 89th percentile of the SAT math test (see figure above).

Quotable:

Sam Altman, CEO of OpenAI, said GPT-4 is more creative than its predecessor, shows less bias and hallucinates a lot less.

It passed a mock bar exam and scored a 5 on some AP exams, he wrote.

However, GPT-4 is „still buggy, still limited, and it still seems more impressive the first time you use it than if you’ve spent more time with it,“ according to Altman.

GPT-4 is also limited when it comes to current events, as most of its training data predates September 2021.