OpenAI announced the latest language model GPT-4, which is an upgraded version of the previous generation GPT-3. The new model has a variety of new capabilities and use cases, including processing words and pictures, better handling of language games, processing more text, and taking exams.
How capable is GPT-4
OpenAI’s latest GPT-4 model is a large multimodal model that accepts image and text inputs and outputs text results. According to OpenAI officials, although it is worse than humans in many real-world situations, it has demonstrated comparable performance to humans in various professional and academic tests.
For example, it scored in the top 10% on the simulated bar exam, while GPT-3.5 scored in the bottom 10%. It took OpenAI 6 months to iteratively adjust GPT-4 through its adversarial testing plan and the lessons of ChatGPT, thus achieving the best results so far (although there is still much room for improvement). Best in terms of maneuverability and refusal to go over the limit.
Over the past two years, OpenAI has rebuilt its entire deep learning stack and, together with Azure, designed a supercomputer designed from the ground up for its workload, officials said. A year ago, OpenAI trained GPT-3.5 as the system’s first “test run.” They found some mistakes and improved the theoretical basis. As a result, their GPT-4 training run was (at least for them) unprecedentedly stable, becoming their first large model to accurately predict training performance.
Features of GPT-4
1. Processing text and images
GPT-4 is a multimodal model that can parse images and text. This means that GPT-4 can analyze the content of the image and connect that information to the written question. Such a function can be applied in many ways. For example, in a demonstration, GPT-4 watched the picture of the inside of the refrigerator, and then was asked what dishes can be made with these ingredients. GPT-4 can give several suggestions based on the image.
2. Better coping with language games
GPT-4 is better at tasks that require creativity or advanced reasoning. In the company’s demo, OpenAI co-founder Greg Brockman asked GPT-4 to summarize a section of an article using only words beginning with the letter “g,” and GPT-4 answered with an almost perfectly correct sentence. GPT-3 doesn’t follow prompts.
3. Handle more characters and types
GPT-4 can process entire scientific papers and novels, allowing it to answer more complex questions and connect more details in any given query. And can remember more contextual information. GPT-4 can handle an input of up to 32,768 Tokens, equivalent to 50 pages of text. This allows GPT-4 to handle hours or days of content computation.
4. Prepare for exams
GPT-4 performs well on a variety of standardized tests, including BAR, LSAT, GRE, and more.
5. Better reasoning skills
GPT-4 is more creative and collaborative than ever. It can generate, edit and work with the user on creative and technical writing tasks such as composing a song, writing a script or learning the user’s writing style. And even better reasoning skills.
6. Stronger role-playing
Features such as tone, style, and personality can be adjusted according to the needs and preferences of the user. GPT-4 provides more parameters and options, allowing users to control the output profile of GPT-4 to suit different scenarios and purposes.
Differences between GPT-3.5 and GPT-4
The difference between GPT-3.5 and GPT-4 is not always noticeable. GPT-4 is more reliable, more creative than GPT-3.5, and able to handle more fine-grained instructions than GPT-3.5.
Key Tips for ChatGPT Operation
However, even though ChatGPT is currently very powerful, it still cannot be regarded as a “sufficiently mature” AI tool, and there are still many parts that must be careful in use.
Although ChatGPT can understand natural language, it does not mean that it can understand “nonsense”. When giving instructions or asking questions, the operator must first ensure that what he says is logical, that is, express his intentions to ChatGPT through clear and clear words and sentences.
Secondly, setting sufficient and appropriate constraints for ChatGPT will definitely help the operator to achieve the desired response results. For example, if you want AI to create a short article, you have to add words such as “within 150 words” to the conversation, forcing ChatGPT to expand or simplify the article.
Accuracy is subject to debate
Even though ChatGPT is smart enough to respond to almost any question a user asks, it doesn’t mean that all the answers it gives are correct. During the actual experience of ChatGPT, the author more or less found that AI gave wrong or plausible answers, and even often chose to shift and blur the focus on unanswerable questions.
The OpenAI team also honestly stated to the outside world that the current ChatGPT has a “disadvantage” of being stubborn and over-guessing the user’s intentions. If the operator insists on arguing with AI about the right or wrong of a thing, they want to educate ChatGPT to answer the correct direction With the answer, although there is a chance to “convince” it, it is more often exhausted by the stubborn AI.
Therefore, the operator should be skeptical about any answer given by ChatGPT. It is definitely an excellent tool with potential, but it needs wise people to use it so that it will not become a fallacy machine and affect the dissemination of correct information.
ChatGPT practical questions and answers
In order for readers to quickly understand the capabilities of ChatGPT, here are a few selected functions for demonstration. Through the text communication process of one question and one answer, we will demonstrate why ChatGPT can cause such a high degree of discussion, and even further help us apply it in our daily life workflow.
1: Responding to General Commonsense Questions
ChatGPT operates as a chatbot, as long as you ask a question through text, AI will come up with its answer. The author first asks ChatGPT with simple questions that are less difficult and can be answered through search engine queries. These two questions are “Where is the capital of Japan?” and “What are the factors of 36?”.
ChatGPT correctly answered that the capital of Japan is Tokyo. The author then asked about the famous tourist attractions in Tokyo.
For the answer “What are the factors of 36?”, ChatGPT not only gave the correct answer, but also explained the definition of factors and how to find them.
2: Answering professional field questions
The most basic general knowledge questions are “a piece of cake” for ChatGPT. Not only are the answers correct and the semantics are clear, but can ChatGPT handle professional questions? In this regard, the author also chose two questions, including “The influence of “The Wealth of Nations” on modern economic theory” and asked ChatGPT to give an answer.
3: Writing Creative Long and Short Articles
The reason why the outside world thinks that ChatGPT is very powerful is that it can not only answer questions, but can even use “creative” writing in plain text. In addition to giving creative themes, we can even specify the style, style and number of words, etc., to make the AI output more in line with the expected results through restrictions. The author hereby gives test to ChatGPT, “Advertising Scripts for Mobile Phones“.
Facing the challenge of creating advertising scripts, ChatGPT’s answers and words are a bit “old-fashioned”, but the clear camera instructions and scene descriptions have successfully allowed the author to imagine the advertising screen in his head. Modifications, definitely enough to be a viable script.
4: Automatically write code
One of the author’s favorite ChatGPT capabilities is probably handing over to AI to write code. Different from answering common-sense and professional questions, if ChatGPT mentions the code in the answer, not only the layout will be changed, which is convenient for users to identify, even the functional annotations will be “filled up”, The format is neat and beautiful, very pleasing to the eye. Next, the author will ask ChatGPT to write a “random password generator” through Python, and then convert it into a “C++” language version.
ChatGPT prepared a “random password generator” written in Python for the author in just 30 seconds. When it was actually executed in the compiler, not only the function was completely normal, but also no errors occurred.
Then I asked ChatGPT to directly convert the previous Python program into a C++ version. After compilation, the function is also normal, and the annotations are clear and clear, and the structure looks beautiful. In the whole process, the author did not write a single line of code.
5: Multilingual AI translation tool
Today’s mainstream online translation tools have actually more or less introduced AI for assistance, so as to obtain more accurate results. ChatGPT, as a large-scale language model, is naturally a great multilingual translation tool, regardless of the word , short sentences or long texts, it is almost impossible to trouble it, but occasionally you may encounter languages that ChatGPT does not support, it is worth noting.
Translate a French paragraph into English through ChatGPT. If you compare the results of other online translation tools, the content and meaning are roughly correct, but the semantic fluency needs to be strengthened. It seems that ChatGPT still has room for improvement in translation tasks.
ChatGPT Login Link
Upgrade ChatGPT
You can use it through OpenAI’s API, but you must be a member of ChatGPT+ (ChatGPT Plus). The current fee is $20 per month, which includes the function of allowing users to use GPT-4. Currently, the number of questions that can be asked per 4 hours is 100.
Basic commands of ChatGPT
1. Use it as a dictionary and look up definitions
- What is “”?
- “”Definition?
Example sentences: “What is half-hearted?”
2. Provide explanations and explain the principles
- How does “” work?
- Please help me explain “”.
Example sentences: “How do corporate mergers and acquisitions work?” “Can you explain Moore’s Law for me?”
3. Advise as a consultant
- What would you suggest for [question/issue]?
- Can you advise me?
4. Provide recommendations:
- Can you recommend a [product/service/solution] for [problem/issue/goal]?
- Should I start with [problem/issue/objective] or [product/service/solution]?
5. Summary information
- Can you do a summary for me?
- What is the main idea of [article/text/news]?
6. Collecting information and citing sources
- Please send me “” articles related to “”
- Help me list news articles related to “” in the past five years, at least “”
7. Provide examples
Can you provide an example of a [concept/process/thing]?
What is a good example of [concept/process/thing]?
8. Translate articles
- Please use “language type” to translate the above article
Conclusion!
In fact, after reading so many instructions and usage examples, you can find that ChatGPT is actually used in a variety of ways, and almost as long as you say it, he can help you achieve it.
Therefore, in fact, the focus is on creativity. You can look at other people’s usage examples on the Internet, see how they say questions, and understand what ChatGPT can do. Reading more and memorizing more can improve your questioning skills.
Through this way of combining punches and combining different commands, your ChatGPT can be ever-changing and help you make more creative things.
