Today OpenAI announced the launch of ChatGPT-4, which ChatGPT Plus users can experience today.
OpenAI calls it “a state-of-the-art system that produces safer and more useful responses.”
Compared with the previous generation, GPT-4 has a wider range of knowledge and stronger problem-solving ability, and performs better on creativity, visual input the long content.
Used on creative projects, it can help users create songs together, write scripts or learn users’ writing styles.
In dealing with long content, GPT-4 can handle texts up to 25,000 words long. Throw a link to it and let it analyze and extract from the text.
It is worth noting that GPT-4 can directly read 32,000 tokens, which is equivalent to giving it 25,000 English words of background information, and can quickly draw conclusions—that is, occupations like paralegals are really going to be replaced up. While the previous version could only read 4096 tokens, equivalent to 3000 English words, the progress of this version is leaps and bounds.
In addition to text, you can also give pictures as input to GPT-4, and it can not only recognize the objects in the picture, but also further process the content based on this information.
According to OpenAI, GPT-4 demonstrates “human-level performance across a variety of professional and academic metrics”.
Similar to the running score software for mobile phones, when ChatGPT first came out, many people used it to “score the test”. Under the GPT-3.5 version model, its SAT score can only rank in the bottom 10%. However, The GPT-4 model can surpass the level of 90% of the test takers.
Similar comparative effects are also shown in other exams. If it is just “running points”, ChatGPT-4 is a level that can be admitted to Harvard and Stanford.
Since he is so good at the exam, can our students use his ability to help us revise and get admitted to a prestigious school? Of course, cheating with it is absolutely not allowed.
What’s more, Open AI also says that GPT 4 is more secure than previous models.
In OpenAI’s internal tests, GPT 4 was 82% less likely to be “tricked” into responding with inappropriate answers, and its response compliance increased by 29% when talking about sensitive topics, such as medical care.
Even so, OpenAI acknowledged that GPT-4 has limitations on “social bias, hallucinations, and adversarial cues.”
OpenAI release timeline recap:
- July 2022: Release of DALL·E
- August 2022: 66% reduction in API fees
- November 2022: Launch ChatGPT
- December 2022: 500% reduction in embedded fees while maintaining state-of-the-art (SoTA)
- March 2023: Release ChatGPT API while maintaining SoTA and reducing fees by 10x
- March 2023: Released Whisper API Released Today: ChatGPT 4
In addition to the guiding use cases that appeared in the official demo, there are also many early adopters on social media who are playing ChatGPT-4, and the cases they give are more instructive for actual application scenarios.
Twitter user @skirano used ChatGPT-4 to code a game program of table tennis “Pong” in less than a minute.
In the era of the computing power explosion, things that we once regarded as classics will be ruthlessly crushed by technological vehicles.
