Four months after ChatGPT’s extremely successful launch, OpenAI is back with an upgraded language model. GPT-4 launches today, and it will soon be integrated into apps like Duolingo and Khan Academy. But there’s a decent chance that you’ve already tested GPT-4.
In a new blog post, Microsoft says that its Bing chatbot already utilizes GPT-4. In fact, it launched with GPT-4. This comes as a bit of a surprise—Microsoft previously described the Bing chatbot as a “more powerful” and “customized” version of GPT-3.5. (To be perfectly clear, Bing’s AI chatbot isn’t a barebones version of GPT-4. The basic GPT-4 model can’t access information beyond September of 2021, among other differences.)
The debut of GPT-4 is actually quite modest. This appears to be an incremental upgrade, and OpenAI CEO Sam Altman warns that GPT-4 “seems more impressive on first use than it does after you spend more time with it.” Based on our experience with the Bing chatbot, I’m inclined to agree. Like GPT-3.5, the new GPT-4 model is prone to inaccuracies and hallucinations.
That said, GPT-4 is still a notable advancement. OpenAI claims that this new language model is “40% more likely” to produce factual responses than its predecessor. And interestingly, it’s 82% less likely to provide “disallowed content” to the user. (But I’m a bit skeptical, as we didn’t have much trouble pulling the initial prompt and other “disallowed content” from Bing’s chatbot.)
And GPT-4 is a multi-modal system, meaning that it can work with more than just text. But this multi-modal system is a bit rudimentary. GPT-4 can “see” images, which should prove to be incredibly useful, but it cannot produce images (barring ASCII images). The model’s output is still limited to language.
I should reiterate that this is an incremental update. GPT-4 still lacks general intelligence, and in CEO Sam Altman’s words, “it is still flawed, still limited.” Overstating or overhyping the capabilities of this AI is a bad idea—language models like GPT-4 are useful and interesting, but they should not be treated as a source of authority, especially in environments like Bing or Google Search.
You can test GPT-4 today with ChatGPT Plus (with a cap of 100 messages every 4 hours). Or, simply join the Bing chatbot early access program. Multi-modal functionality isn’t working at this time; presumably, it will launch after the GPT-4 developer event on March 14th at 4PM ET.
Source: OpenAI
