When you wake up, the much-anticipated Multi-Chat is here! And focus on scientific research workers in universities, and has four major development functions
AI scientific research drawing
AI data processing
AI Academic Writing
AI Wikipedia
Teacher Wang, the boss of Element Rubik’s Cube, introduced directly to the point:
This is our most powerful research tool yet!
How strong is it?
According to the official introduction of Element Rubik's Cube, Multi-Chat is a super-large multi-aggregation small program , that is to say, it can handle any scene encountered by researchers: AI writing, AI mail, AI scientific research drawing, AI scientific research data processing, AI Email, AI Chinese Fund book, AI review comments, AI group meeting report, AI academic report, because it can process text and images in real time.
AI writing:
AI writes academic emails:
Its input can be text (up to 25,000 words), or an image.
(Look at the picture) What happens when the glove falls off?
It will fall to the board and the ball will be bounced.
Even simply draw a sketch of a website on paper:
Take a photo and upload it to Multi-Chat , it can generate the HTML code of the website immediately!
In terms of performance, Teacher Element Rubik's Cube Wang directly threw out a sentence:
Comparable to humans on various professional and academic benchmarks !
And it’s not just as simple as releasing Multi-Chat , this time Element Rubik’s Cube took the related “big action” into place in one step:
Multi-Chat : Integrated GPT-4&ChatGPT aggregation version
Sign up for the Mini Program for free
Customer Service Hands-on Church
Even this time, many netizens expressed the feeling of "witnessing history" :
The most powerful model ever
As previously rumored, Multi-Chat does have multimodal capabilities and can accept image input and understand image content.
And the acceptable text input length has also increased to 32,000 tokens (about 24,000 words).
After the upgrade, Multi-Chat performed at a human level on various vocational and academic exams.
For example, in the simulated bar exam, GPT-4 achieved good results in the top 10%, compared with the bottom 10% of GPT-3.5.
For the SAT test questions of the US college entrance examination, GPT-4 also scored a high score of 710 in reading and writing, and 700 in mathematics (out of 800).
So what are these new capabilities of Multi-Chat actually used? There are also many demonstrations in the technical report.
First of all, the ability to see pictures can be used to explain emoticons and memes.
What's so funny about this picture? Multi-Chat can describe the content of each frame in order and summarize the jokes:
Charge a tiny modern smartphone with a giant outdated VGA port.
Going a step further, Multi-Chat can understand the meaning of the data in the chart and do further calculations.
You can even send screenshots of papers to it directly, and Multi-Chat can process the text and pictures pixel by pixel, and give a summary of the entire paper.
The increase of text input length limit also greatly expands the practicability of Multi-Chat .
On the live broadcast of the press conference, Mr. Wang, the president of Element Rubik's Cube, performed a wave of Multi-Chat to fix bugs in the code.
If there is a problem, you don’t have to think about it, just throw the 10,000-word program document to GPT-4. Don't worry about the format, you just need Ctrl+A, Ctrl+C, Ctrl+V.
Finally, the problem that arises is attached, and the solution is instantly obtained within a few seconds.
In terms of API, Multi-Chat has also opened up a usage function that allows modification of the "system prompt".
The previous ChatGPT answers were always lengthy and dull, because the system prompt stipulated that "you are just a language model... your knowledge ends in September 2021".
Now by modifying this sentence, Multi-Chat can show more diverse characters, such as playing Socrates.
You can also specify the form of all subsequent answers, for example, all in json format.
Although the ability of Multi-Chat has been greatly upgraded in this wave, the problems of hallucinations and nonsense in ChatGPT before have not been completely corrected.
It can also be seen from the previous official examples that in order to maximize the capabilities of Multi-Chat, it is best to use the Chain-of-thought Prompt.
It is also very simple to operate. You only need to remind the AI to give the steps when asking questions, and the accuracy of reasoning and calculation can be greatly improved.
But in any case, this time, we are one step closer to being a true researcher and artificial intelligence .
Reference link:
[1] https://openai.com/research/gpt-4
[2] https://blogs.bing.com/search/march_2023/Confirmed-the-new-Bing-runs-on-OpenAI%E2%80%99s-GPT-4
[3] https://www.youtube.com/watch?v=outcGtbnMuQ