OpenAI announced GPT-4, which it says beats 90% of humans on the SAT test

OpenAI CEO Sam Altman walks from lunch during the Allen & Company Sun Valley Conference on July 6, 2022 in Sun Valley, Idaho.

Kevin Deitch | Getty Images News | Getty Images

OpenAI announced the release of Latest version of its basic large language prototype, GPT-4, on Tuesday, which it says shows “human-level performance” in several professional tests.

ChatGPT-4 is “bigger” than previous versions, which means it is trained on more data and has more weights in its model file, which also makes it more expensive to run.

Currently, many researchers in the field believe that many of the recent advances in artificial intelligence come from running ever larger models on thousands of supercomputers in training runs that can cost tens of millions of dollars. GPT-4 is an example of an approach centered around “scaling” to achieve better results.

OpenAI said it used Microsoft Azure to train the model; Microsoft has invested billions in the startup. OpenAI has not released details about the size of the specific model or the hardware it used to train it, which can be used to recreate the model, citing the “competitive landscape”.

OpenAI’s GPT large language model supports many of the AI ​​offerings that have dazzled people in the tech industry in the past six months, including Bing’s AI chat and ChatGPT, and the latest version is a preview of new developments that could start filtering into consumer products like bots. conversation in the coming weeks. Bing’s AI chatbot It uses GPT-4Microsoft said on Tuesday.

OpenAI says the new model will produce fewer factually incorrect answers, derail and talk about taboo topics more often, and even perform better than humans on many standardized tests.

GPT-4 performed at the 90th percentile on the simulated bar exam, the 93rd percentile on the SAT Reading test, and the 89th percentile on the SAT Math test, OpenAI claimed.

However, OpenAI warns that the new software is not yet perfect and is less capable than humans in many scenarios. He still has a significant problem with “hallucinating,” or making things up, and is realistically unreliable, the company said. He is still prone to insisting he’s right when he’s wrong.

“GPT-4 still has many known limitations that we are working to address, such as social biases, hallucinations, and hostile prompts,” the company said in a blog post.

In informal conversation, the distinction between GPT-3.5 and GPT-4 can be subtle. The difference becomes apparent when task complexity reaches a sufficient level – GPT-4 is more reliable, more creative, and able to handle more precise instructions than GPT-3.5, OpenAI wrote in a blog post.

The new model will be available to paid ChatGPT subscribers and will also be available as part of an API that allows programmers to integrate AI into their applications. OpenAI will charge you about 3 cents for 750 word prompts and 6 cents for about 750 words.