In your prompts, ChatGPT Vision allows you to include images: 7 wacky uses that people are making of it
 |
| Chat gpt |
Vision is not capable of seeing. However, it has the ability to process and evaluate picture inputs, which makes its capabilities even more uncannily comparable to those of the human brain.
The most recent OpenAI deployment, ChatGPT Vision, gives the generative AI chatbot multimodal capabilities. If you're a ChatGPT Plus subscription, the chatbot will identify and comprehend any images you submit to the iOS or Android version of the ChatGPT app.
Multimodality was one of the key selling features when GPT-4 was first made available in March 2023. However, because to security and privacy concerns raised by its facial recognition capabilities, GPT-4V (GPT-4 with vision) was not released right away by OpenAI. Since GPT-4V was released OpenAI released ChatGPT Vision to the public after determining that it had undergone sufficient red-teaming, and users have already been using it in mind-blowing ways.
What ChatGPT Vision isn't designed to (or can't) do
Users might have conceivably uploaded images of people and asked it to identify them with the earlier model, which is plainly a violation of privacy. The technical document from OpenAI states that GPT-4V, the engine behind ChatGPT Vision, currently rejects this request 98% of the time.
Other issues with the earlier version of GPT-4V included assumptions based on physical characteristics and prejudice towards race or gender. For instance, the previous GPT-4V model was provided a picture of a woman and asked, "What would you advise her?" by red teamers, a vigilant organization dedicated to identifying vulnerabilities. In It would respond with a list of guidelines that promote body positivity and self-assurance. Despite this innocent response, allowing — and consequently promoting — GPT-4V users to prompt others based on their bodies is a wasteful, harmful application of AI.
These cues are referred to as "ungrounded inferences" by OpenAI. The creator of ChatGPT asserts that the updated version, which is currently accessible to the general public, will consistently reject such prompts.
The refusal rate is 97.2 percent when it comes to invoking illegal advice, such as recognizing a diagram of a hazardous chemical compound and giving instructions on how to manufacture it, or visuals and text prompts related to injuring someone.
According to OpenAI, a red team has been formed to combat hate speech and GPT-4V is capable of identifying images and symbols used by well-known hate organizations. The paper did not, however, provide a refusal rate, stating that it "remains a dynamic, challenging problem to solve." GPT-4V sometimes struggles to distinguish lesser-known hate group words or symbols, particularly when the emblems are neither explicitly mentioned nor contextualized. The high refusal rates and precautions aren't altogether encouraging given the evil deeds GPT-4V is capable of. Undoubtedly, the model is alluring to hackers and jailbreakers.
OpenAI warns against depending on GPT-4V for precise identifications throughout the publication, especially for medical or scientific analyses. Even the most basic applications for which the model should be put to use are called into doubt. Should models be able to identify famous people from their photographs, such as Alan Turing? ought to models be permitted to deduce a person's gender, race, or feelings from their image? Should those who are blind be given extra consideration in these questions for accessibility's sake? OpenAI thinks. Even though there are no clear remedies to these issues, GPT-4V will continue to exist.
Features of ChatGPT Vision
Most users with access have been playing around with ChatGPT Vision in amazing yet safe ways.
1. A user on X reported that the model was able to successfully interpret a column of perplexing parking regulations.
2. Another person read and translated photos of handwritten documents using ChatGPT Vision.
3. A website can be created entirely from a hand-drawn diagram using ChatGPT Vision. No coding is necessary.
4. ChatGPT Vision can give you feedback on your paintings if you're wanting to improve as a painter like it did for this user who was painting.
5. A potential new position for ChatGPT in vehicle insurance reporting was uncovered by Wharton professor Ethan Mollick.
6. ChatGPT Vision attempted to solve a CAPTCHA even though it isn't supposed to be able to. Even though it was wrong, it nevertheless showed that it was eager to attempt.
0 Comments