OpenAI launched GPT-4 on March 14 and says the updated version is more reliable and creative than its predecessor.
CEO Sam Altman said the tech was capable of passing the bar exam and "could score a 5 on several AP exams."
ChatGPT was previously powered by GPT-3.5. The update is available to users who pay for ChatGPT Plus and access to the API will be granted to a limited number of developers on OpenAI's waitlist.
But those who have tried Microsoft's artificial-intelligence-powered Bing have already experienced some of the new tech. On Tuesday, Microsoft said its AI-boosted Bing had been powered by a version of GPT-4 that was "customized for search."
Here are five key ways the update differs from GPT-3.5.
1. GPT-4 can understand images
GPT-4 is "multimodal," meaning it can see and process image prompts as well as text.
Users can ask the chatbot to describe images, but it can also contextualize and understand them. In one example given by OpenAI, the chatbot is shown describing what's funny about a group of images.
The chatbot is still limited to text responses and cannot produce images itself.
2. The bot is said to be more accurate
According to OpenAI, the update will give more-accurate responses to users' queries.
OpenAI said in a blog post that the system was "40% more likely to produce factual responses than GPT-3.5." GPT-4 also has more "advanced reasoning capabilities" than its predecessor, according to the company.
The updated chatbot is still not immune to "hallucinations," a tendency for AI to generate false responses or reasoning errors. OpenAI said the chatbot was not perfect. Altman called it "still flawed, still limited."
3. Users can have longer conversations
GPT-4 can take in and generate about eight times more text than ChatGPT.
This means that the chatbot will have a longer "memory" and be able to keep up with lengthier conversations. OpenAI said the latest version could process up to 25,000 words, compared with the previous 3,000 words.
4. It's harder to break the rules
GPT-4 may be bad news for fans of ChatGPT's "evil" alter egos.
OpenAI said the update included its best-ever results on "steerability, and refusing to go outside of guardrails."
The company said that the system was "82% less likely to respond to requests for disallowed content." Many users have attempted to trick ChatGPT into answering inappropriately or overriding its content moderation, likely providing OpenAI with many examples of malicious prompts.
5. The chatbot is more creative
OpenAI said the update was the most creative and collaborative version of ChatGPT yet.
The company said the changes may be "subtle" in casual conversations but would become clear when the bot's faced with complex situations. GPT-4 is also "able to handle much more nuanced instructions than GPT-3.5," OpenAI said.
In collaboration with users, the chatbot can produce and edit creative-writing tasks such as drafting screenplays. The company added that the updated chatbot could learn a user's writing style.
