Artificial intelligence research lab OpenAI has released GPT-4, the latest version of the groundbreaking AI system that powers ChatGPT, which it says is more creative, less prone to fabricating facts, and less biased than its predecessor.
Calling it “our most capable and aligned model yet,” OpenAI co-founder Sam Altman said the new system is a “multimodal” model, meaning it can accept images and text as inputs, allowing users to users ask questions about the images. The new version can handle massive text input and can remember and act on more than 20,000 words at a time, allowing you to take an entire novel as a message.
The new model is available today for users of ChatGPT Plus, the paid version of the ChatGPT chatbot, which provided some of the training data for the latest version.
OpenAI has also worked with commercial partners to offer services powered by GPT-4. A new subscription level of the Duolingo language learning app, Duolingo Max, will now offer English-speaking users AI-powered conversations in French or Spanish, and can use GPT-4 to explain mistakes students have made in languages. On the other end of the spectrum, payment processing company Stripe uses GPT-4 to answer support questions from corporate users and help spot potential scammers on the company’s support forums.
“Artificial intelligence has always been a big part of our strategy,” said Duolingo’s senior product manager, Edwin Bodge. “We had been using it to customize lessons and run Duolingo English tests. But there were gaps in the learning journey that we wanted to fill: conversational practice and contextual feedback on mistakes.” The company’s experiments with GPT-4 convinced it that the technology was capable of providing those features, with “95% ” of the prototype created in one day.
During a demo of GPT-4 on Tuesday, Open AI president and co-founder Greg Brockman also gave users a sneak peek at the image recognition capabilities of the newest version of the system, which is not yet publicly available. and it’s only being tested by a company called Be My Eyes. The feature will allow GPT-4 to analyze and respond to images that are sent along with prompts and answer questions or perform tasks based on those images. “GPT-4 is not just a language model, it’s also a vision model,” Brockman said, “it can flexibly accept input that intersperses images and text in an arbitrary way, like a document.”
At one point in the demo, GPT-4 was asked to describe why the image of a squirrel with a camera was funny. (Because “we don’t expect them to use a camera or act like humans.”) At another point, Brockman sent a photo of a crude, hand-drawn sketch of a website to GPT-4, and the system created a website based on it. drawing.
OpenAI claims that GPT-4 fixes or improves many of the criticisms that users had with the previous version of their system. As a “big language model”, GPT-4 is trained on vast amounts of data pulled from the Internet and attempts to provide answers to sentences and questions that are statistically similar to those that already exist in the real world. But that may mean making up information when you don’t know the exact answer, a problem known as “hallucination,” or providing annoying or abusive answers when given the wrong directions.
By leveraging the conversations users were having with ChatGPT, OpenAI says it was able to improve, but not eliminate, those weaknesses in GPT-4, sensitively responding to requests for content, such as self-harm or medical advice, “29% more often ” and answering incorrectly. to disallowed content requests 82% less frequently.
However, GPT-4 will still “blow up” the facts, and OpenAI cautions users: “Extreme care should be taken when using language model output, particularly in high-risk contexts, with the exact protocol (such as human review , grounding with additional context, or avoiding high-risk uses altogether) that match the needs of a specific use case.” But it scores “40% higher” in tests meant to measure hallucinations, says OpenAI.
The system is particularly good at avoiding cliché: older versions of GPT will gleefully insist that the statement “you can’t teach an old dog new tricks” is factually accurate, but newer GPT-4 will correctly tell the user who asks if you can teach an old dog new tricks that “yes you can.”