AI Competition: Can GPT4 and ChatGPT help us create better Chinese websites?

​Author | Yun Zhao

It was initially said to be released on Thursday, but it turned out that it was released in the early morning of Wednesday. Sure enough, OpenAI always likes to surprise people in advance!

#1. GPT4, ChatGPT Competition

Without further ado, let’s start with the hard skills of GPT4.

First of all, it is the ability to process long text. This is described on the OpenAI official website: "GPT4 can process text of more than 25,000 words, allowing the use of long-form content to create, expand conversations, and Use cases such as document search and analysis."

OpenAI is afraid that everyone can't understand it vividly, so we simply used a long article from Wikipedia as an example. How long is this text? The editor actually clicked on the link (see the picture below), which is twice as long as the text that ChatGPT (GPT3.5) can currently handle. Friends who don’t believe it can go to this link: https://en.wikipedia.org/wiki/Rihanna to take a look at the actual length.

For text dialogue tasks, how do the two perform? OpenAI official said this.

In casual conversation, the difference between GPT3.5 and GPT4 can be subtle. When the complexity of the task reaches a sufficient threshold, differences become apparent. GPT4 is more reliable, more creative, and capable of handling more nuanced instructions than GPT3.5.

The context length of GPT4 is 8192 tokens. OpenAI also provides limited access to 32768 context (~50 pages of text) version gpt-4-32k, which will also be automatically updated over time (the current version, gpt-4-22k-0314, is also supported until June 14th). Pricing is $0.06 per 1K prompt tokens and $0.12 per 1K completion tokens.

OpenAI is still improving model quality in the long-term environment and would like feedback on how it performs in use cases. OpenAI handles requests for the 8K and 32K engines at different rates based on capacity, so they can be accessed at different times.

Second, the ability to recognize images: accept images as input.

GPT4 can accept prompts for both text and images, in parallel with the text-only setting, allowing the user to specify any visual or language task. Specifically, it generates textual output (natural language, code, etc.) given an input consisting of interspersed text and images. GPT4 exhibits similar functionality to plain text input in a range of areas, including documents with text and photos, diagrams or screenshots. Additionally, it can be enhanced with test-time techniques developed for text-only language models, including a handful of shots and thought chain prompts. Unfortunately, image input is still only in the research preview stage and has not been made public for the time being.

(Questions and answers are all in English: for ease of understanding, they have been translated into Chinese)

Some ingredients are given in the picture. Ask GPT4 what kind of delicacies can be made based on the ingredients in the picture. It can be seen that GPT4 has officially crossed the border!

The third one is more reliable reasoning ability.

GPT4 is more creative and collaborative than previous models. Not only can it complete a lot of creative and basic writing, but it can also create songs, scripts, and importantly, learn the user's writing style.

Regarding reasoning capabilities, OpenAI also gave an example of letting GPT4 reserve a conference room:

The problem is roughly this: three people have different free time every day, let GPT find the appropriate time for a 30-minute meeting.

It can be seen that ChatGPT’s logical ideas and text analysis are all correct, but the answers given are all wrong. GPT4 gave the correct answer.

GPT4 also has a stronger BUFF. If it takes the exam, it will crush many top students! And ChatGPT is so weak!

achieved the top 10% on the Uniform Bar Examination, while ChatGPT received the bottom 10%.

In addition, OpenAI also gives rankings for other exams such as SAT, which are better than ChatGPT!

Compared with ChatGPT, what other hard goods are there? ​

GPT4’s operability is more flexible. Friends who have used ChatGPT know that if you submit a request using the API, there are generally two messages: system message (specified AI style and role) and user message (specific question raised). The previous ChatGPT did not pay much attention to system messages, and the style and tone were relatively fixed.

This is not the case with GPT4, where developers (and soon ChatGPT users) can now dictate their AI style and tasks by describing these directions in a "system" message. Allow API users to significantly customize the user experience within scope. OpenAI will continue to make improvements here (in particular, be aware that system messages are the easiest way to "jailbreak" the current model, i.e., respect for boundaries is no longer so rigid), and OpenAI is also very encouraged to let them know if users want to try these ideas .

It also has stronger support for multiple languages. Among the 26 languages ​​tested, 24 are more accurate than GPT3.5 and other language models.

The editor opened his eyes and looked around, but he didn’t see any Chinese characters. I decisively went to find out how effective the relevant Chinese test was, and the result was disappointing. Not only is it slow, but it also gives the wrong answer!

Is there any place where ChatGPT is ahead of GPT4?

have! That's how fast text is generated, GPT4 generates answers significantly slower than ChatGPT. It seems that the more capable you are, the more cautious you will be in answering questions!

2. Actual evaluation

The first example is: a sketch on a napkin can make GPT4 magically generate a web page image!

Source: Twitter

Real example, a user on Twitter posted a After taking a photo of the sketch on the napkin, I handed it over to GPT4, and it turned into a fully functional html/css/javascript website.

Netizens exclaimed: This model can actually read and interpret anything written on a napkin!

Another netizen had an idea: Could it be an application based on GPT4 that turns my written prescription into something that pharmacists can actually read, and it can become the next one? American biotech billionaire.

Second example: Rebuilding the game with GPT4.

A user named Pietro recreated the Pong game in 60 seconds, and the score is pretty good.

Photo source: Twitter

Photo source: Twitter

GPT4 can solve puzzles more accurately due to its extensive general knowledge and problem-solving abilities. The creativity of GPT4 is so terrifying.

3. Limitations of GPT4

Although GPT4 has stronger capabilities than ChatGPT, its limitations are still similar to the early GPT model. Most importantly, it's still not completely reliable (it still makes up facts and makes errors of reasoning). When using language model output, especially in high-risk contexts, great care should be taken to use precise protocols that match the needs of the specific use case (such as human review, additional context bases, or avoiding high-risk use entirely).

In OpenAI’s internal adversarial authenticity evaluation, GPT4 scores 40% higher than our latest GPT3.5. In other words, even if the answer generated by GPT4 is more accurate, it is only 40% more credible than GPT4!

4. Access permissions

ChatGPT Plus users can obtain GPT4 access permissions from chat.openai.com. In the meantime, OpenAI will adjust the exact usage cap based on actual demand and system performance, but OpenAI expects severe capacity constraints (although OpenAI will be expanded and optimized in the coming months).

Depending on the traffic patterns OpenAI sees, OpenAI may introduce new subscription levels for higher volume GPT4 usage; OpenAI also hopes to offer some free GPT4 queries at some point , so those without a subscription can also try it.

To access the GPT-4 API (which uses the same ChatCompletions API as GPT-3.5-turbo), you also need to sign up for OpenAI's waitlist.

OpenAI will invite some developers starting today and gradually expand the scale to balance capacity and demand. If you are a researcher with social impact, you can also apply for use through OpenAI’s researcher access program.

5. The application ecosystem has been explored

OpenAI uses GPT4 internally, which has had a huge impact on functions such as support, sales, content management, and programming. OpenAI also uses it to help humans evaluate AI output, beginning the second phase of OpenAI’s alignment strategy.

OpenAI looks forward to GPT4 becoming a valuable tool to improve people’s lives by powering many applications. There is still much work to be done, and OpenAI looks forward to improving, exploring, and contributing to this model through the collective efforts of community building. The editor noticed that OpenAI has reached a cooperation with a virtual volunteer tool company.

Partners of Be My Eyes. Connecting people in need of vision support with volunteers and residents and companies through factual videos from around the world.

Simplify tasks for people with low vision

With the Be My Eyes app Programs that support visually impaired people and help them achieve independence. Blind or low vision users may need support including checking expiration dates, differentiating colors, reading instructions, or navigating new environments.

6. Written at the end

OpenAI calls GPT4 the latest milestone in its efforts to expand deep learning. As a large-scale multi-modal model (accepting image and text input, emitting text output), GPT4 demonstrates human-level performance on a variety of professional and academic benchmarks.

Although there is a big gap with our previous multi-modal capability predictions (text, image, audio, video and other inputs), we can see the evolution of the GPT model It is making solid progress, especially in terms of understanding general knowledge and handling more difficult problems. I am really amazed!

The above is the detailed content of AI Competition: Can GPT4 and ChatGPT help us create better Chinese websites?. For more information, please follow other related articles on the PHP Chinese website!