'More creative with higher accuracy': Say hello to OpenAI's latest, GPT-4
After much hype, OpenAI unveiled ChatGPT's successor, GPT-4, last night. A large multimodal model, GPT-4, can process image and text inputs, though it can only respond through text. The company claims the model is "more creative and collaborative than ever before" and "can solve difficult problems with greater accuracy".
In a very impressive demo, Greg Brockman, president and cofounder of OpenAI, showcased GPT-4 and some of its capabilities and limitations.
Brockman showed the technology doing taxes, calculating a married couple's standard deduction and total tax liability. "This model is so good at mental math. It has these broad capabilities that are so flexible," Brockman said in the demonstration video.
Announcing GPT-4, a large multimodal model, with our best-ever results on capabilities and alignment: https://t.co/TwLFssyALF pic.twitter.com/lYWwPjZbSg— OpenAI (@OpenAI) March 14, 2023
GPT-4 can 'see', 'read', and 'reason'
According to a blog post, while the upgraded version of deep learning is less capable than humans in many real-world scenarios, it "exhibits human-level performance" on various professional and academic benchmarks.
This, it proved through a simulated bar exam. Ever since ChatGPT's release, academics and researchers have been putting it to the test, and the chatbot has been performing well. GPT-4 passed the bar exam with a score in the top 10 percent of test takers. This contrasts with GPT-3.5's score, which was in the bottom 10 percent. GPT-4 was also able to score a five on several AP exams.
A technical report on the language model clearly states that GPT-4 "outperforms both previous large language models and most state-of-the-art systems".
OpenAI says that it spent six months pouring over its latest language model, ensuring it was safer and more aligned. Lessons were taken from ChatGPT, and their adversarial testing program, which has resulted in their "best-ever results".
Perhaps, one of the most interesting aspects of GPT-4 is that it can 'see'.
GPT-4 can "reason" based on a prompt of images which lets the user specify any vision or language task. It generates text outputs when given inputs that consist of images and texts. OpenAI said that image inputs are still a research preview and not publicly available.
OpenAI has already partnered with companies to incorporate GPT-4 into their products
While the new model is available via ChatGPT Plus, which requires a monthly ChatGPT subscription for $20, it will also be accessible as an API for developers to build on, replete with a waitlist.
OpenAI has already collaborated with companies to integrate GPT-4 into their products. Popular language and learning app Duolingo introduced Duolingo Max, powered by GPT-4. Payment procession platform Stripe and education platform Khan Academy will leverage the enhanced capabilities of GPT-4 for a pleasing user experience. Microsoft has also revealed that GPT-4 has been powering their Bing chatbot all this time.
Interestingly, OpenAI has not unveiled the size of GPT-4 or how it was built. "OpenAI is now a fully closed company with scientific communication akin to press releases for products," Thomas Wolf, cofounder of Hugging Face, the AI startup behind the open-source large language model BLOOM, told MIT Technology Review.
However, like former language models, OpenAI warns that GPT-4 isn't completely reliable and could "hallucinate". GPT-4 also lacks knowledge about events "that have occurred after the vast majority of its data cuts off" in September 2021.
Do animals break up in the same way that we do? Do they consider it breaking up at all?