SOPA Images via Getty Images
OpenAI has announced a new generation of its GPT engine called GPT-4. GPT stands for Generative Pre-trained Transformer, and compared to the current GPT-3.5, GPT-4's main improvement is its ability to accept image inputs in addition to text inputs, although its output remains text-based. In addition to model improvements, GPT-4 has also incorporated user feedback from ChatGPT over the past six months to further improve output results.
OpenAI explains that in simpler questions and conversations, GPT-4's performance will not differ much from GPT-3, but in more complex questions, GPT-4 excels in the appropriateness, correctness, and rejection of crossing the set "guardrails" compared to GPT-3. As for GPT-3.5, it is actually the same basic model as GPT-3, but was developed in collaboration with Azure to create a supercomputer specifically designed for GPT model training, along with accompanying software. OpenAI was able to complete the training of GPT-4 quickly and stably by gaining experience from GPT-3.5.
OpenAI plans to make GPT-4 available to the public through updates to ChatGPT and APIs. However, if you have successfully secured a spot to use Bing AI, it has actually been using the GPT-4 model since it was made available for public testing, instead of ChatGPT's GPT-3.5. However, since Bing's usage differs slightly from ChatGPT, relying more on curated search network information rather than generating text based on prompts, it is difficult to directly compare the two. If you are interested, you can join the waiting queue to try out Bing AI.
As for image input, it is currently still in the research preview phase, with no open plans yet. OpenAI provided some usage examples in the article, such as describing the content of an image, analyzing charts, answering exam questions that include images, identifying unusual aspects of an image, and explaining riddles and jokes, which looks very impressive.
The biggest risk for ChatGPT and similar AIs is that they may spout nonsense with a straight face. In this regard, GPT-4 has reduced the probability of generating "fantasy" content by around 40% compared to its predecessor, but it is still not zero. Therefore, OpenAI strongly recommends considering the appropriateness of using products and services based on GPT-4 based on usage needs. For high-importance information, it is best not to rely solely on the content provided by GPT-4 or use human verification to ensure accuracy.