Shin Ji Won reports
Edit: Peach
Mathematical genius Tao Zhexuan's research also uses the AI tool GPT-4. With GPT-4, all you want to do, just copy, paste, and do it with one click!
When the red fried chicken ChatGPT has also become a research tool for mathematical genius Tao Zhexuan.
Recently, he said online that he found some small use cases for ChatGPT.
First, it's pretty good at parsing documents in code format (in this case the API for #arXiv searches) and then returning a properly formatted code query (later it also provides some working python code to call this API the way I asked it to, even though I had to manually install a package to get it to run). Second, I asked it to come up with questions that smart students might ask in an undergraduate linear algebra class (for which I provided some sample questions), and it gave some great examples that inspired me on the possible direction of the course, as well as potential homework questions.
In summary, I found that AI tools like ChatGPT, while not helping me solve an unsolved math problem, were useful for other work-related problems, although some manual adjustments were often required afterwards.
There is little difference between GPT-4 and GPT-3.5
How good is GPT-4 at parsing documents in code format?
Tao had GPT-4 search for all arXiv articles whose metadata was a given keyword (e.g., "Collatz") and submitted within a certain date range (e.g., between 2011-2016). According to the following documentation, what is the format of such a query?
In this regard, GPT-4 not only gives the code, but also lists the classification of the query.
Netizens asked, as a reading tool, given a paper, let it ask what questions a reviewer or speaker might ask? How well does GPT-4 do in this regard?
In this regard, Tao Zhexuan took it to evaluate and praised it as "really good".
"Here are the first few pages of a recent paper. Can you ask some questions that the audience might ask while listening to a lecture about this paper?"
ChatGPT replied, "Of course, here are some questions that mathematicians might ask after listening to a lecture on this paper." Balabala then lists 7 questions.
Careful netizens found that the screenshot posted by Tao Zhexuan was GPT-4.
"What do you think is the difference between the old ChatGPT and the new assistant based on the GPT-4 model?"
"So far, GPT-4's response has been slightly more useful and meets my requirements, but I have not seen a significant difference in quality," Tao said.
ChatGPT joins the workflow
With the birth of ChatGPT, many people are exclaiming that their jobs will be replaced by AI.
Even the genius mathematician Tao Zhexuan got started with ChatGPT.
On March 5, Tao Zhexuan said online that he had decided to try to incorporate AI tools into my workflow in a different way. These include GPT-4, DeepL and other tools.
During this time, he dug up ChatGPT's hidden functions, such as finding formulas, parsing documents in code format, rewriting paper sentences, and so on.
For example, ChatGPT can sometimes do a semi-finished semantic search in mathematics, that is, use it to generate some prompts.
For example, Tao Zhexuan asked ChatGPT to identify Kummer's theorem from the description, which failed to give the correct answer, but according to the approximate answer it gave (the Legendre formula).
In this regard, Tao Zhexuan said that the role of artificial intelligence in mathematics is to provide a preliminary approximate answer, and then it can be combined with traditional search engines to easily find the correct answer.
Tao also found ChatGPT a bright spot when dealing with mathematical problems, being able to recognize transliterated versions of mathematical concepts in different languages.
Alternatively, ChatGPT can be asked to convert a bunch of references obtained from MathSciNet and format it as \bibitems in the context of the LaTeX bibliography.
But isn't ChatGPT ever wrong?
In a question about the proof of "whether there are infinite primes", Tao found that the answer given by ChatGPT was not entirely correct.
On the other hand, he found that the argument given by ChatGPT could be fixed, and this idea he had never seen before.
Mining the use cases of GPT-4, netizens have also spent a lot of effort.
Having GPT-4 review papers is not only reasonable, but also hits the point.
"As an academic, I don't know how to look at this: I put an old paper of mine in GPT-4 (split in two) and asked an economic sociologist for rigorous but fair peer review."
It creates a perfectly sound peer review that hits many of the points raised by my reviewers.
For many people, what do they want to do, just ask GPT-4, "Copy, paste, one-click done!"
It's like making a complex game with GPT-4.
Netizens formed GPT-4, Replit, MidJourney, and Claude into an artificial intelligence assistant team and created a 3D space parkour game from scratch.
But Goose doesn't know anything about JavaScript or 3D game programming.
Resources:
https://mathstodon.xyz/@tao