ChatGPT creator OpenAI debuts new GPT-4 AI system

OpenAI said Morgan Stanley is using GPT-4 to organize data, while Stripe Inc., an electronic payments company, is testing whether it will help combat fraud. Other customers include language learning company Duolingo Inc., the Khan Academy and the Icelandic government.

Be My Eyes, a company that works on tools for people who are blind or have low vision, is also using the software for a virtual volunteer service that lets people send images to an AI-powered service, which will answer questions and provide visual assistance.

“We’re really starting to get to systems that are actually quite capable and can give you new ideas and help you understand things that you couldn’t otherwise,” said Greg Brockman, president and co-founder of OpenAI.

The new version is better at things like finding specific information in a corporate earnings report, or providing an answer about a detailed part of the US federal tax code — basically combing through “dense business legalese” to find an answer, he said.

Like GPT-3, GPT-4 can’t reason about current events — it was trained on data that, for the most part, existed before September 2021.

The earlier version of the ChatGPT chat screen on a smartphone.

In a January interview, OpenAI Chief Executive Officer Sam Altman tried to keep expectations in check.

“The GPT-4 rumor mill is a ridiculous thing,” he said. “I don’t know where it all comes from. People are begging to be disappointed and they will be.” The company’s chief technology officer, Mira Murati, told Fast Company earlier this month that “less hype would be good.”

GPT-4 is what’s called a large language model, a type of AI system that analyzes vast quantities of writing from across the internet in order to determine how to generate human-sounding text. The technology has spurred excitement as well as controversy in recent months. In addition to fears that text-generation systems will be used to cheat on schoolwork, it can perpetuate biases and misinformation.

When OpenAI initially released GPT-2 in 2019, it opted to make only part of the model public because of concerns about malicious use. Researchers have noted that large language models can sometimes meander off topic or wade into inappropriate or racist speech. They’ve also raised concerns about the carbon emissions associated with all the computing power needed to train and run these AI models.

Read more about the environmental implications of AI’s computing power

OpenAI said it spent six months making the artificial intelligence software safer. For example, the final version of GPT-4 is better at handling questions about how to create a bomb or where to buy cheap cigarettes — for the latter case, it now offers a warning about the health impacts of smoking along with possible ways to save money on tobacco products.

“GPT-4 still has many known limitations that we are working to address, such as social biases, hallucinations and adversarial prompts,” the company said Tuesday in a blog, referring to things like submitting a prompt or question designed to provoke an unfavorable action or damage the system. “We encourage and facilitate transparency, user education and wider AI literacy as society adopts these models. We also aim to expand the avenues of input people have in shaping our models.”

The company declined to provide specific technical information about GPT-4 including the size of the model. Brockman, the company’s president, said OpenAI expects cutting-edge models will be developed in the future by companies spending on billion-dollar supercomputers and some of the most advanced tools will come with risks. OpenAI wants to keep some parts of their work secret to give the startup “some breathing room to really focus on safety and get it right.”

It’s an approach that is controversial in the AI field. Some other companies and experts say safety will be improved by more openness and making the artificial intelligence models available publicly. OpenAI also said that while it is keeping some details of model training confidential, it is providing more information on what it’s doing to root out bias and make the product more responsible.

“We have actually been very transparent about the safety training stage,” said Sandhini Agarwal, an OpenAI policy researcher.

The release is part of a flood of AI announcements coming from OpenAI and backer Microsoft, as well as rivals in the nascent industry. Companies have released new chatbots, AI-powered search and novel ways to embed the technology in corporate software meant for salespeople and office workers. GPT-4, like OpenAI’s other recent models, was trained on Microsoft’s Azure cloud platform.

Google-backed Anthropic, a startup founded by former OpenAI executives, announced the release of its Claude chatbot to business customers earlier Tuesday.

Alphabet Inc.’s Google, meanwhile, said it is giving customers access to some of its language models, and Microsoft is scheduled to talk Thursday about how it plans to offer AI features for Office software.

The flurry of new general-purpose AI models is also raising questions about the copyright and ownership, both when the AI programs create something that looks similar to existing content and around whether these systems should be able to use other people’s art, writing and programming code to train. Lawsuits have been filed against OpenAI, Microsoft and rivals.