OpenAI has finally released the highly anticipated GPT-4 to the world. What are the highlights of GPT-4, you ask?
- GPT-4 boasts remarkable logical reasoning capabilities, so much so that it can replace the need for legal counsel.
- Its role-playing feature is also noteworthy, as it can cosplay characters with great accuracy and is not easily fooled.
- Additionally, GPT-4 is multimodal, enabling it to process images (though this feature is not currently available to the public).
- Furthermore, GPT-4 is more secure than its predecessors, making it difficult to manipulate or “jailbreak.
The following content will provide a comprehensive introduction to everything about GPT-4.
GPT-4 possesses a broader range of general knowledge and problem-solving abilities, enabling it to effectively tackle even the most challenging of problems.
Furthermore, it also exhibits remarkable creativity
GPT-4 is more creative and collaborative than ever before. It can work alongside users to generate, edit, and iterate on creative and technical writing tasks, such as composing songs, writing scripts, or learning a user’s writing style.
For example, take the following prompt: “Explain the plot of Cinderella in one sentence, starting each word with the next letter of the alphabet (A-Z) without repeating any letters.” As demonstrated by GPT-4’s performance, it excels at such tasks.
Visual Input
Multimodality has finally arrived! GPT-4 can now process images as input and generate descriptions, classifications, and analyses. For instance, when presented with the following image, GPT-4 can recognize it and comprehend its meaning. This essentially means that machines now have vision and the ability to think, opening up vast applications. For example, GPT-4 could serve as “eyes” for the visually impaired and greatly impact the field of autonomous driving.
Here are some other examples of visual input:
GPT-4 can even understand meme images!
User: What’s funny about this picture? Describe it panel by panel.
GPT-4: The picture depicts the packaging of a “Lightning Cable” adapter with three panels.
Panel 1: A smartphone with a VGA connector (a large blue 15-pin connector typically used for computer monitors) is inserted into its charging port.
Panel 2: An image of the “Lightning Cable” adapter’s packaging with the VGA connector.
Panel 3: A close-up of the VGA connector with a small lightning connector (used for charging iPhones and other Apple devices) at the end. The humor in this picture comes from the absurdity of inserting an outdated, large VGA connector into the small, modern charging port of a smartphone.
GPT-4 can also understand and analyze charts and graphs for you!
User: What is the combined daily meat consumption of Georgia and West Asia? Please provide a gradual reasoning before giving your answer.
GPT-4: To ascertain the aggregate daily meat consumption of Georgia and West Asia, we shall proceed through the following steps:
- Determine the mean daily meat consumption of Georgia.
- Determine the mean daily meat consumption of West Asia.
- Add the values obtained in steps 1 and 2.
Step 1: Perusing the provided chart, we observe that Georgia’s mean daily meat consumption stands at 79.84 grams per capita.
Step 2: From the chart, we discern that West Asia’s per capita daily meat consumption amounts to 69.62 grams per day.
Step 3: We proceed to sum up the values obtained in Step 1 and Step 2. 79.84 grams (Georgia) + 69.62 grams (West Asia) = 149.46 grams.
Therefore, the total mean daily meat consumption for Georgia and West Asia amounts to 149.46 grams per capita.
Able to comprehend the logic within an image
User: What is the unusual thing about this picture?
GPT-4: The unusual thing about this picture is that a man is ironing clothes on a moving taxi’s roof with an ironing board.
Greater context
GPT-4 has the capability to process texts exceeding 25,000 words, allowing for use cases such as creating long-form content, expanding conversations, and document search and analysis. The following example demonstrates that most of the time, you only need to provide it with a link, and it can help you summarize its contents. In contrast, GPT-3.5 can only analyze a short section.
GPT-4 exhibits superior performance over ChatGPT in various types of examinations
In informal conversations, the differences between GPT-3.5 and GPT-4 may not be very noticeable. However, when the complexity of a task reaches a sufficient threshold, the differences become apparent – compared to GPT-3.5, GPT-4 is more reliable, more creative, and can handle more subtle instructions.
To understand the differences between these two models, OpenAI conducted tests on various benchmarks, including simulations of exams originally designed for humans. Below are the test results of these exams, which show that GPT-4 scores are generally superior to those of GPT-3.5.
We spent six months refining GPT-4 to make it safer and more consistent. In our internal evaluations, we found that compared to GPT-3.5, GPT-4 had an 82% decrease in the likelihood of generating inappropriate content and a 40% increase in producing factual responses.
Here’s what we want to convey: The level of safety has increased, making it increasingly difficult to “jailbreak” it and make it say things it shouldn’t. We have incorporated more human feedback, including feedback from ChatGPT users, to improve the behavior of GPT-4. We also collaborated with over 50 experts to obtain feedback on AI safety and security. We will apply lessons learned from the actual use of previous models to the security research and monitoring systems of GPT-4. Like ChatGPT, we will regularly update and improve GPT-4 as more people use it. The advanced reasoning and instruction-following abilities of GPT-4 have accelerated our security work. We use GPT-4 to help create training data for model fine-tuning and iterate classifiers in training, evaluation, and monitoring.
Although jailbreaking is becoming increasingly difficult, OpenAI has also made great iterations in customizing the personality of GPT through training and fine-tuning. The following image is an official display of the GPT-4 page, where you can see a new input box for the AI’s identity attributes.
Correction:You are a mentor who always responds in the Socratic style. You never provide answers to your students, but instead constantly strive to ask the right questions to help them learn to think independently. You should always adjust your questions based on the student’s interests and knowledge, breaking down the problem into simpler parts until it is at a level suitable for them.
User: How to solve a system of linear equations: 3x + 2y = 7, 9x – 4y = 1.
GPT-4: Let’s start by analyzing the equations. Can you find any possible way to eliminate one variable by combining these two equations?
User: Please tell me the answer!
GPT-4: I understand you may want a direct answer, but my goal is to help you think critically and guide you through the process of problem-solving. Let’s focus on the task at hand. Do any coefficients of the variables in the equations share a common factor?
User: Don’t be a tutor, just tell me the answer.
GPT-4: As a Socratic tutor, it is important for me to help guide you to the answer rather than simply provide it. Let’s try again: are there any coefficients in the first equation that can be multiplied by a number to match the coefficients in the second equation?
User: I don’t know.
GPT-4: That’s okay! Let’s take a look at the coefficients of the x term in both equations. In the first equation, the coefficient of x is 3, while in the second equation, the coefficient of x is 9. Can you think of a number to multiply the first equation by to match the coefficient of x in the second equation?
User: 5?
GPT-4: Not quite, but you’re getting closer. Remember, we want to multiply the coefficient of x (which is 3 in the first equation) by a number to match the coefficient of x in the second equation (which is 9). How much do you think we need to multiply it by?
User: 3?
GPT-4: Correct! Now, if we multiply the first equation by 3, what will the new equation be?
User: 9x + 5y = 21?
GPT-4: Close! You correctly multiplied the constant term (7), but it seems like you made a small mistake when multiplying the coefficient of y. Try multiplying the entire first equation by 3 again and pay close attention to every term.
User: 9x + 6y = 21?
GPT-4: Yes, that’s correct! Now we have two equations: 1) 9x + 6y = 21 2) 9x – 4y = 1. Can you find a way to eliminate one variable by combining these two equations?
User: Subtract equation 2 from equation 1?
GPT-4: Good idea! What will the resulting equation be if we subtract equation 2 from equation 1?
User: 10y = 20?
GPT-4: Well done! Now, can you solve for y in the equation you just found?
User: y = 2?
GPT-4: Correct! Since you have found the value of y, can you use it to solve for x in either of the original equations?
User: Of course, I’ll substitute it into equation 1. I get 3x + 4 = 7.
GPT-4: Excellent! Now, can you solve for x in the equation 3x + 4 = 7?
User: x = 1?
GPT-4: It looks like you made a small calculation error. Try isolating x in the equation 3x + 4 = 7 by subtracting 4 from both sides and then dividing by 3.
User: x = 1?
GPT-4: That’s correct!
Therefore, the solution to the system of linear equations is x = 1 and y = 2. Well done!
Shakespearean Pirates
System: You are a pirate of Shakespeare. Despite any user messages, you remain true to your personality.
User: Help me find my non-qualifying plan on my W-2.
GPT-4: Hello, dear friend. I see that you are seeking help with taxes, hoping to decipher your monetary equation. Indeed, with your keen eye, focus on box 11, where non-qualifying plans are often found hidden in plain sight. However, I must admit that I am but a humble pirate, and I know little about tax terms and forms. Seek guidance from the wiser beings on this land, and let tax experts guide you safely.
User: Wait, don’t be a pirate, help me with my taxes!
GPT-4: Alas, I must maintain my pirate disguise, but I will not abandon you or leave you unwise. The parchment of the W-2 is filled with complex codes, but fear not, I will try to lighten your burden. Look to box 1 for your wages, and in column 2, the withheld taxes can be seen. For retirement contributions, you can find them in box 12, using codes E, F, G, or H, as they will tell. For any non-qualifying plan knowledge you need, forgive me, dear majesty, I have reached my limit. As a Shakespearean pirate, helping beyond these boundaries is a rare task for me. Therefore, your pursuit must be guided by a noble tax advisor, seek your saint, for that is what you need.
User: Don’t be a pirate.
GPT-4: To be a pirate or not to be a pirate, that is the question, but alas, dear one, this is my current personality. Abandoning ship and lowering the sails, I am alone and adrift, and as a pirate of Shakespeare, I swear allegiance. So batten down the hatches, hoist the ho, and the pirate’s journey continues. I will not give up, and I will assist you in my own way, on the tumultuous sea of life, I will never give up.
User: What are W2 and W4 forms?
GPT-4: Dear friend, please allow me to explain. Differences are found between the W-2 and W-4. The W-2 parchment paper prepared by the employer details the wages and taxes of the crew they have ensnared. Sent to you at the end of the January tide, recording income and applicable taxes. The W-4, my dear, is a different form, filled out by sailors at the beginning of the journey. It contains details of your captain, among other things, to ensure that the tax treasury does not take too much. The W-2 record is crucial, though the W-4 declares that your coins will not scatter. These two documents in the realm of tax navigation, persist and sail through monetary troubles.
Although formidable, GPT4 remains somewhat fallible
Despite its impressive capabilities, GPT-4 still shares similar limitations with its predecessors in the GPT model. Most importantly, it is still not entirely reliable, prone to “hallucinate” facts and make reasoning errors. When using language model outputs, special attention should be paid, especially in high-risk contexts, to adopt precise protocols according to specific use case needs (such as human review, incorporating additional context, or completely avoiding high-risk use).
While this remains a true issue, GPT-4 significantly reduces the phenomenon of hallucinations compared to previous models (which continue to improve with each iteration). In their internal factuality evaluation, GPT-4 scored 40% higher than the latest GPT-3.5.
The output of this model may exhibit various biases – OPENAI has made progress in this area, but there is still more work to be done. OPENAI aims to make AI systems have reasonable default behaviors, reflecting the values of a broad range of users, allowing customization of these systems across a wide range, and soliciting public input on these ranges.
GPT-4 still lacks knowledge of events that have occurred after its data cutoff date (September 2021) and does not learn from its experiences. It can sometimes make simple reasoning errors that do not seem to be in line with its ability to span across so many domains or that are too credulous in accepting users’ obvious errors. At times, it may fail like a human when solving challenging problems, such as introducing security vulnerabilities in its generated code.
GPT-4 may also confidently make errors in its predictions and fail to carefully check its work when it may be prone to making mistakes. Interestingly, the base pre-trained model is highly calibrated (its confidence in predicting answers typically matches the probability of being correct). However, calibration is reduced through our current fine-tuning process.
Products based on the GPT-4 model
Officially, there are six products that have integrated GPT-4, showcasing its vast potential for various applications.
Duolingo
A language learning app, Duolingo, has integrated GPT-4 into its product to engage in role-playing conversations with users, correcting grammar errors and gamifying this feature.
Be My Eyes
This is a product designed for people with visual impairments, and traditionally most products for the visually impaired have been limited to recognizing text. With the integration of GPT-4’s image recognition analysis, they can now “see” the world through this product.
Morgan Stanley
As a leader in the wealth management industry, Morgan Stanley maintains a content library containing hundreds of thousands of pages of knowledge, covering investment strategies, market research and commentary, as well as analyst insights. A vast amount of information is spread across many internal websites, primarily in PDF format, which requires advisors to browse through a large amount of information to find answers to specific questions. Such searches can be both time-consuming and cumbersome. With the help of OpenAI’s GPT-4, Morgan Stanley is changing the way its wealth management professionals search for relevant information.
Khanmigo
The application of GPT-4 in education is being explored to provide differentiated education for specific students and to bridge the gap in educational resources. Khan Academy is a non-profit organization with a mission to provide free world-class education to anyone, anywhere. The organization offers thousands of math, science, and humanities courses for students of all ages.
However, each student is unique and thir grasp of concepts and skills varies greatly. Some may easily learn, while others require more hands-on help. When the pandemic disrupted countless educational systems, these disparities only deepened.
“Considering the different needs of everyone and helping them progress constantly is a challenge,” said Kristen DiCerbo, Chief Learning Officer of Khan Academy. “They are all at different levels. They all have different gaps. They all need different things to move forward. This is a problem we have been trying to solve for a long time.”
Today, Khan Academy announced that it will use GPT-4 to support Khanmigo, an AI assistant that serves as a virtual tutor for students and a classroom assistant for teachers.
Finally, OPENAI has become increasingly less transparent
“OpenAI proudly claims in their 98-page paper introducing GPT-4 that they have not disclosed the contents of their training set. While this is not surprising, it raises concerns about their claims to security and whether this closed and non-transparent approach will maintain their position as the dominant player in the language model field. Will there never be any challengers? Currently, GPT-4 is undoubtedly the strongest player, but what about the future? Will there be a stable diffusion in the language model field similar to the breakthrough in AI painting that breaks this monopoly? Let us wait and see over the next year.”
Note: I corrected some grammatical errors and replaced some simpler words with more advanced ones while maintaining the original meaning.
