Just months after ChatGPT stunned technology heavyweights like Google, OpenAI announced its latest and most advanced language model. The new GPT-4 model can solve complex problems with greater accuracy than the old model, but it’s also safer, with a lower chance of being tricked into saying something evil. OpenAI says it has already partnered with several companies to build GPT-4 into products, including one you’ve heard a lot about: Bing. The “next-generation” model Microsoft bragged about in its new Bing bot was GPT-4 all along.
Like the old GTP-3.5, GPT-4 is a type of language model known as a transformer. These algorithms were invented several years ago at Google, but OpenAI was the first to turn them into a real product. A “generative pre-trained transformer” needs a lot of data in the pre-training phase, which teaches the AI to predict the next word in a passage. This forms the basis for how the model generates new text from a smaller data set when you ask it to do your homework (which you shouldn’t do).
Many online have speculated that GPT-4 would represent a huge leap in AI capabilities, but it’s more of an iterative update. According to OpenAI, GPT-4 has advanced reasoning capabilities, which makes the output more natural and similar to what a human might work out. Much to the potential chagrin of educators, GPT-4 is also an even more capable test-taker. While GPT-3.5 could perform well enough on tests like the SATs and LSAT, GPT-4 beats most humans. It scores in the 90th percentile of the Uniform Bar Exam, compared with the 10th percentile for GPT-3.5.
The new AI-powered Duolingo Max service promises a lifelike conversational partner to help you learn. Credit: OpenAI/Duolingo
OpenAI is honest about the limitations of GPT-4. While it’s 40% more likely to produce factual responses and 82% less likely to respond to “disallowed content,” it can still exhibit social bias, aggressiveness, and AI hallucinations. That’s a nice way of saying it gets things wrong while believing that it’s right. Remember when the Bing bot refused to accept that it was 2023 and not 2022? OpenAI confirms that kind of thing can still happen, but it’s working on solutions.
The ChatGPT you can access on OpenAI’s site is based on the 3.5 model. However, Microsoft has confirmed that the Bing chatbot has been running GPT-4 all along. Previously, the company only said it was using a newer version of OpenAI’s technology than you could get elsewhere. This technology is no longer limited to Bing, though. OpenAI has worked with companies like Duolingo, Stripe, and Khan Academy to build GPT-4 into their products. You can try the new GPT-4 model with OpenAI’s $20 monthly ChatGPT Plus subscription, and there will be an API for developers soon. Interested parties can sign up to be notified when the service opens up.