Tao Zhexuan announced that he would chair the White House Generative AI Working Group, and Li Feifei and Hassabis gave speeches

Recently, the U.S. President’s Council of Advisors on Science and Technology (PCAST) established a working group on generative artificial intelligence.

It is worth mentioning that the mathematical genius Terence Tao plays the role of co-leader in this working group.

Tao Zhexuan posted on his blog that Laura Greene and I co-chair this generative artificial intelligence working group.

His blog stated that this group mainly studies the broader impact of generative artificial intelligence technology in science and society. Impact, including popular text-based large language models (such as ChatGPT), diffusion models for image generation (such as DALL-E2, Midjourney), and scientific application models (such as protein design or weather forecasting).

Tao Zhexuan chairs the generative AI group

The White House mentioned in an article published on the 13th that PCAST ​​established the generative AI The group helps assess key AI opportunities and risks and provides input on how best to ensure the development and deployment of these technologies is as fair, responsible and safe as possible.

At the end of the article, you can see that among the members of the working group, Terence Tao is among them.

In addition, AMD CEO Lisa Su is also a member of this generative AI group member.

According to Terence Tao’s blog, the generative AI group will be held during the PCAST ​​conference on Friday, May 19 Hold a public meeting.

Live link: https://www.whitehouse.gov/pcast/meetings/2023-meetings/

Two expert panels will present the current state of generative AI, concluding with an extensive Q&A session. These speakers include:

Artificial Intelligence for Science:

##Anima Anandkumar (Caltech & Nvidia)

Demis Hassabis (Google DeepMind)

Li Feifei (Stanford)

Artificial Intelligence & Society:

Sendhil Mullainathan (Chicago)

Daron Acemoglu (MIT)

Sarah Kreps (Cornell University)

In addition, Tao also mentioned that the working group is working on how to promote the active deployment of generative artificial intelligence applications, and how to best Reduce risks well and seek public input.

The initial focus is on how to detect, resist, and mitigate false information and DeepFake generated by artificial intelligence without sacrificing freedom of speech. This is a challenging topic.

After ChatGPT was added to the workflow

After ChatGPT came out, it also gained the favor of mathematics experts like Terence Tao.

In his latest post on Mathstodon, he shared his views on generative AI tools.

I started to see the comparative advantages between myself and current generative AI tools. I already have enough skills to optimize the tasks I perform every day, so AI tools don’t really help me much. Most obviously in studying math, but also writing emails.

For those tasks for which I have some expertise but little practice, AI tools are helpful: often I can use them to create a first draft of the output, which I can then validate and modified, or at least used as a source of inspiration. Sometimes I will find that the shortcomings of AI may inspire me. Although this is also in line with the essence of Cunningham's Law, it is still more efficient to use AI than to seek answers independently. Examples of this type include processing data, translating into foreign languages, and writing in formats that I rarely use, such as public speaking, rules documents, etc.

For those tasks for which I have little expertise and do not require extremely high quality and reliable output, one can simply ask the AI ​​tool and more or less follow its lead suggestion. Here, artificial intelligence functions as a slightly more convenient version of traditional search engines.

Finally, for those tasks for which I have no expertise but require quality and reliability that neither AI nor myself can solve, I must consult human experts. Such as repairing a complex, expensive, delicate piece of equipment.

In short, Tao Zhexuan’s views on generative AI tools are divided into four categories. It is worth noting that there is not much added value in his field of mathematics expertise.

For the third situation, Tao gave an example by asking ChatGPT to summarize the previous article into a flow chart.

ChatGPT provides a description of the text. Tao said that it is speculated that future multi-modal GPT will be able to directly provide a flow chart instead of giving a text description.

An example for the second category of tasks: After realizing that I could ask GPT to output the flowchart in LaTeX format, I got the first image below, which is obviously not perfect of. But since I'm familiar with LaTeX, it's not difficult to correct it manually to the second image.

Obviously, Tao has also used ChatPDF in the workflow, which became popular some time ago.

#In March, Tao said that he decided to try to incorporate AI tools into his workflow in different ways. These include tools such as ChatGPT and DeepL.

#In the following period, he often shared some of his experiences using tools such as ChatGPT.

Many hidden functions of ChatGPT were discovered by him, such as finding formulas, parsing documents in code format, rewriting thesis statements, etc.

For example, ChatGPT can sometimes do a semi-finished semantic retrieval in mathematics, that is, use it to generate some hints.

For example, Terence Tao asked ChatGPT to identify Kummer's theorem from the description. It failed to give the correct answer, but based on the approximate answer it gave (Legendre's 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 Zhexuan also discovered the highlight of ChatGPT in dealing with mathematical problems, being able to identify transliterated versions of mathematical concepts in different languages.

Alternatively, ChatGPT can be asked to convert a bunch of references obtained from MathSciNet and put them in a LaTeX bibliographic environment Formatted as \bibitems.

But doesn’t ChatGPT make mistakes?

In a question about proving whether there are infinitely many prime numbers, Terence Tao discovered that the answer given by ChatGPT was not completely correct.

On the other hand, he found that the argumentation ideas given by ChatGPT could be fixed, and he had never seen this idea before.

Have AI tools such as ChatGPT been added to your workflow?

The above is the detailed content of Tao Zhexuan announced that he would chair the White House Generative AI Working Group, and Li Feifei and Hassabis gave speeches. For more information, please follow other related articles on the PHP Chinese website!