In an unexpected announcement today, OpenAI released the long-awaited GPT-4 model, an update of the technology behind its popular chatbot, ChatGPT. The company is calling GPT-4 its “most advanced system, producing safer and more useful responses.” The surprise announcement comes less than four months after ChatGPT debuted and became the fastest-growing consumer application in history.
GPT-4 advances the core technology of ChatGPT by enabling the chat software to solve more difficult problems with greater accuracy, thanks to its broader general knowledge and problem solving abilities. It also adds new capabilities such as accepting images as inputs and generating captions, classifications, and analyses. GPT-4 is also capable of handling over 25,000 words of text, allowing for use cases like long-form content creation, extended conversations, and document search and analysis.
The announcement came as a surprise to many in the tech community, which widely believed GPT-4 would be announced at Microsoft’s Future of Work with AI event on Thursday. The news dropped on the same day that Google announced a laundry list of new generative AI capabilities, including a PaLM API and new features in Google Cloud and Google Workspace.
One of the biggest updates in the GPT-4 model is that it introduces the first elements of multimodal chat, meaning it can work with more modalities than just text. GPT-4 can accept images as inputs and generate captions, classifications, and analyses. GPT-3.5, its predecessor, could only accept text. While GPT-4’s capabilities fall short of text-to-video generation and other dynamic generative content, it does offer a glimpse of what a multimodal chat will look like in the future. It’s easy to imagine how, in the future, video, audio, images, and all other forms of content will be integrated into the chat.
OpenAI says GPT-4 surpasses ChatGPT
In a company blog post, OpenAI claims that GPT-4 “surpasses ChatGPT in its advanced reasoning capabilities,” and “leverages more data and more computation to create increasingly sophisticated and capable language models. The company adds, “We spent 6 months making GPT-4 safer and more aligned. GPT-4 is 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5 on our internal evaluations.”
In an effort to add greater sophistication to the model, GPT-4 also incorporates more human feedback, including feedback submitted by ChatGPT users, to improve GPT-4’s behavior. OpenAI says the company worked “with over 50 experts for early feedback in domains including AI safety and security.”
As part of the announcement, OpenAI shared several use cases from companies that have already tested GPT-4, including Duolingo, Stripe, Morgan Stanley and the Government of Iceland. The company also shared a GPT-4 research blog and emphasized that “GPT-4 still has many known limitations that we are working to address, such as social biases, hallucinations, and adversarial prompts.”
Suresh Venkatasubramanian, a computer scientist and professor at Brown University, remains cautiously optimistic about the announcement. In an interview with VentureBeat, he said, “I’m eager to see how it behaves under the stress testing that ChatGPT went through in the public gaze, and I’m especially interested in seeing whether and how it hallucinates deceptively real content and what protections are in place to prevent that.”
GPT-4 is available on subscription plan ChatGPT Plus and as an API for developers.
VentureBeat’s mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.
Sharon Goldman
Source link
