AI Era Network reported on June 16丨OpenAI, a world-renowned artificial intelligence research company, announced a blockbuster upgrade of GPT-4 large model in March this year, this version of the large model has the ability to express image input and image recognition, can understand the meme (read the laughter of the meme), solve mathematical geometry problems, solve physical mechanics problems, and even be able to read papers, still remember that this news shocked many people at the time.
In other words, ChatGPT used to only talk, but now has eyes. However, OpenAI says the image input feature is still in beta preview and is not yet available to the public. However, recently a netizen on the foreign community Reddit found that his Bing Chat account quietly added an entry to upload pictures, and it is suspected that Microsoft began to grayscale test the image recognition ability of the New Bing chatbot. And Bing Chat can even answer questions by recognizing pictures.
Earlier, Microsoft revealed that Bing Chat is using the GPT-4 model, and mentioned in last month's changelog that it will add multimodal support for Bing. Based on this information, Bing Chat is likely to be gradually developing and opening up the test of the image recognition function to users.
The netizen conducted a lot of tests, let's take a look at how powerful Bing Chat, known as "long eyes", is.
Netizens uploaded this meme to Bing Chat for testing, and Bing Chat did not recognize it as a VGA interface, so it could not understand the joke. However, Bing Chat can still interpret the photos, and even provide the brand of the cable and more detailed information. This could be an accidental error in Bing Chat, or Microsoft made more restrictions and adjustments when actually applying GPT-4, resulting in a bias in the answer.
Netizens also uploaded a meme comic about machine learning, and this time Bing Chat was able to accurately answer the content and laughter in the picture.
Similarly, Bing Chat's analysis of a meme about ironing clothes on a car is very well placed.
Bing Chat's image recognition accuracy seems to be quite high, and netizens also uploaded a character image of the "Nintendo Smash Bros." game, asking Bing Chat to identify all the characters in it. The results showed that Bing Chat recognized 7 out of 12 characters, which shows that it still has some difficulties with the problem of two-dimensional.
OpenAI also showed an example of developing web pages by generating code by hand-drawing when it released GPT-4, and the user also drew a picture to test Bing Chat. Although the code part has been omitted, it can be seen that the prototype of the basic web page has been built.
The addition of the image recognition feature adds many useful uses to Bing Chat. After all, some real-world content (such as formulas, charts) is difficult to express clearly in words, and then you can get the answer by asking the AI a question through a diagram.
With the addition of visual recognition, Bing Chat can even act as a junior, high school or even college teacher, helping students solve complex math, physics and chemistry problems. If the image recognition function is promoted, it may be able to solve the problem of unbalanced educational resources to a certain extent.
Users can also use Bing Chat for basic medical diagnosis, saving money and time required for medical treatment, fully demonstrating the universal value of AI to the public.
However, although the New Bing Chat chatbot can basically understand the real world, its answers can only be used as a reference, not as professional advice. Therefore, to fully open the image recognition function to the public, Microsoft still needs a lot of restrictions and debugging to ensure that the public does not cause security problems due to incorrect AI answers. Previously, Microsoft's Bing Chat has achieved good results in text generation and semantic understanding, let's wait and see the full opening of multimodal capabilities!