OpenAI unveils GPT-4, months after ChatGPT stunned Silicon Valley

Four months ago, a San Francisco small business became the talk of the tech industry when it introduced a new online chatbot capable of answering complex questions, writing poetry, and even… imitate human emotions.

Now the company is back with a new version of the technology that powers its chatbots. The system will up the ante in Silicon Valley’s race to embrace artificial intelligence and decide who will be the next generation of tech industry leaders.

OpenAI, which has about 375 employees but has been backed by billions of dollars in investment from Microsoft and industry celebrities, said Tuesday it has released technology it calls GPT-4. It was designed to be the underlying engine that powers chatbots and all sorts of other systems, from search engines to personal online tutors.

Most people will use this technology through a new version of the company’s ChatGPT chatbot, while businesses will integrate it into a wide variety of systems, including enterprise software and e-commerce websites. The tech is already piloting the chatbot available to a limited number of people using Microsoft’s Bing search engine.

Advances in OpenAI have, in just a few short months, propelled the tech industry into one of its most unpredictable moments in decades. Many industry leaders believe that developments in AI represent a fundamental technological change, as significant as the creation of web browsers in the early 1990s. The rapid improvement has stunned computer scientists.

GPT-4, which learns its skills by analyzing huge amounts of data mined from the Internet, improves on what powered the original ChatGPT in several ways. It’s more precise. It can, for example, pass the Uniform Bar Exam, instantly calculate a person’s tax debt, and provide detailed image descriptions.

But OpenAI’s new technology still has some of the weirdly human shortcomings that have vexed industry insiders and pissed off people who’ve worked with the latest chatbots. He is an expert on some subjects and a dilettante on others. He can do better on standardized tests than most people and offer accurate medical advice to doctors, but he can also mess up basic arithmetic.

Companies betting their future on technology may – at least for now – have to accept vagueness, which has long been taboo in an industry built from the ground up on the idea that computers are more demanding. than their human creators.

“I don’t want to give the impression that we’ve solved the reasoning or the intelligence, which we certainly haven’t,” OpenAI chief executive Sam Altman said in an interview. “But it’s a big step up from what already exists.”

Other tech companies are likely to include GPT-4 functionality in a range of products and services, including software from Microsoft to perform business tasks and e-commerce sites that want to offer customers new ways to do business. try their products virtually. A number of industry giants like Google and Facebook’s parent company Meta are also working on their own chatbots and AI technology.

ChatGPT and similar technologies are already changing the behavior of students and educators trying to figure out whether the tools should be adopted or banned. Because systems can write computer programs and perform other business tasks, they are also about to change the nature of work.

Even the most impressive systems tend to complement skilled workers rather than replace them. Systems cannot be used in place of doctors, lawyers or accountants. Experts are still needed to spot their mistakes. But they may soon replace some paralegals (whose work is reviewed and edited by trained lawyers), and many AI experts believe they will replace workers who moderate content on the internet.

“There are definitely disruptions, which means some jobs are disappearing and new jobs are being created,” said Greg Brockman, president of OpenAI. “But I think the net effect is that the barriers to entry go down and the productivity of experts goes up.”

On Tuesday, OpenAI began selling access to GPT-4 so businesses and other software developers can build their own apps on it. The company has also used the technology to create a new version of its popular chatbot, which is available to anyone who purchases access to ChatGPT Plus – a subscription service priced at $20 per month.

A handful of companies are already working with GPT-4. Morgan Stanley Wealth Management implements a system that instantly retrieves information from company documents and other records, and delivers it to financial advisors in conversational prose. Khan Academy, an online education company, uses technology to create an automated tutor.

“This new technology can act more like a tutor,” said Khan Academy CEO and Founder Sal Khan. “We want him to teach new skills to the student while the student is doing most of the work.”

Like similar technologies, the new system sometimes “hallucinates”. It generates completely fake information without warning. When asked about websites that feature the latest cancer research, he could give several Internet addresses that don’t exist.

GPT-4 is a neural network, a type of mathematical system that learns skills by analyzing data. This is the same technology that digital assistants like Siri use to recognize voice commands and that self-driving cars use to identify pedestrians.

Around 2018, companies like Google and OpenAI started building neural networks that learned huge amounts of digital text, including books, Wikipedia articles, chat logs, and other information posted on the Internet. . They are called Large Language Models or LLMs.

By identifying billions of patterns in all that text, LLMs learn to generate text themselves, including tweets, poems, and computer programs. OpenAI threw more and more data into its LLM. More data, the company hoped, would mean better answers.

OpenAI has also refined this technology using feedback from human testers. As people tested ChatGPT, they rated the chatbot’s responses, separating those that were helpful and truthful from those that weren’t. Then, using a technique called reinforcement learning, the system spent months analyzing those ratings and better understanding what it should and shouldn’t do.

“Humans rate the things they like to see and the things they don’t like to see,” said OpenAI researcher Luke Metz.

The original ChatGPT was based on a large language model called GPT-3.5. OpenAI’s GPT-4 learned from much larger amounts of data.

OpenAI executives declined to disclose how much data the new chatbot learned from, but Brockman said the dataset was “internet-scale”, meaning that it covered enough websites to provide a representative sample of all English speakers on the Internet.

The new capabilities of GPT-4 may not be obvious to the average person using the technology for the first time. But they’re likely to be developed quickly as laypeople and experts continue to use the service.

Given a long New York Times article and asked to summarize it, the bot will give an accurate summary almost every time. Add a few random sentences to this summary and ask the chatbot if the revised summary is accurate, and it will flag the added sentences as the only inaccuracies.

Mr Altman described the behavior as “reasoning”. But technology cannot duplicate human reasoning. He is good at analyzing, summarizing and answering complex questions about a book or a news article. He is much less skilful if asked about events that have not yet happened.

He can write a joke, but he doesn’t show he understands what will actually make someone laugh. “It doesn’t capture the nuance of what’s funny,” said Oren Etzioni, founding chief executive of the Allen Institute for AI, a premier lab in Seattle.

As with similar technologies, users can find ways to trick the system into strange and creepy behavior. Asked to imitate another person or play, this type of bot sometimes veers into areas it was designed to steer clear of.

GPT-4 can also respond to images. From a photograph, graph or diagram, the technology can provide a detailed description of the image in several paragraphs and answer questions about its content. It could be a useful technology for the visually impaired.

On a recent afternoon, Mr. Brockman demonstrated how the system reacts to images. He gave the new chatbot an image from the Hubble Space Telescope and asked him to describe the photo “in great detail”. He responded with a four-paragraph description, which included an explanation of the ethereal white line that stretched across the photo. A “trail of a satellite or a shooting star,” the chatbot wrote.

OpenAI executives said the company isn’t immediately releasing the image description portion of the technology because they don’t know how it could be misused.

Building and serving chatbots is extremely expensive. Because it’s trained on even greater amounts of data, OpenAI’s new chatbot will increase business costs. Mira Murati, chief technology officer of OpenAI, said the company could restrict access to the service if it generates too much traffic.

But in the long term, OpenAI plans to build and deploy systems that can juggle multiple media types, including sound and video as well as text and images.

“We can take all of these general-purpose knowledge skills and break them down into all sorts of different areas,” Brockman said. “It takes the technology into a whole new realm.”