🪴 Quartz 4.0

Talk about applying for OpenAI's "Democratic inputs to AI".

Jul 21, 20249 min read

for 3 2023-07-05 by Yasukazu Nishio

”Democratic Inputs to AI” in OpenAI

  • Announced on 5/25: Democratic inputs to AI
  • OpenAI “supports experiments to improve democratic processes.”
  • Total $1,000,000.00 … One hundred and ten thousand yen?
  • What is this? democratic process? Why would OpenAI pay 100 million for this?

Objective of “Democratic Input to AI”

.

  • AI has far-reaching economic and social implications

  • It is not right that decision-making of “what AI should be” is done by “a few people” such as the management of a for-profit company.

  • Support experimental projects to take the first steps in this direction

  • Recruit a team to develop a proof of concept of a democratic process that can answer the question, “What are the rules that AI systems should follow?”

  • We (OpenAI) want to learn from these experiments and use them as the basis for a more global, more ambitious process.

  • Sounds really interesting, doesn’t it?/nishio/nishio.icon

    • It would be interesting to participate in “an attempt for the future of AI” that OpenAI is flagging!
    • want to be involved in creating [the basis for a more global, more ambitious process!

Applied!!!!!!!!

  • We’ll know the results on 7/15.
  • Normally we don’t talk about this kind of thing in the undecided stage, but as we started planning the project, the project members themselves started to get excited and said, “It’s interesting, so let’s do it regardless of whether it’s accepted or not.
  • We are already active and accepting applications for participation in the project.

Things you must do if adopted

.

  • requirement
    • Over 500 people will participate in the experiment.
    • A report on the findings from the experiment will be published on 10/20.
  • This requirement of “more than 500 participants” seems difficult…?

Code for Japan

  • Here’s the Code for Japan story
  • Code for Japan is a general incorporated association established in 2014

    • Code for Japan’s Objective: To improve public services through citizen participation processes and contribute to society through better government and local government.

  • Participatory Democracy Platform Decidim in Japan from 2020
  • I applied for this project as a Code for Japan project.
    • Mr. Haruyuki Seki, the president of the board, is one of the members.
      • Five hundred? That won’t be a problem!”
    • Code for Japan will do the contract with OpenAI, etc. as a legal entity.
    • (It would be great if a corporate entity with a track record of activities could act in such cases, as it would make things much easier/nishio/nishio.icon)
  • The feasibility of this process has increased dramatically!

Project Structure

.

  • A little “pizza” parable before going into the details
    • Once you have a “pizza base,” you can add various “toppings” on top of it.
    • The minimal execution of the project is ensured by the pizza’s base, which allows for individual experimentation in toppings.
    • Toppings can be tried and tested as you like, and only the ones that work should be released in the end.
    • I think everyone here is capable of making unique toppings.
  • The foundation of the pizza is the iterative structure of “SF Prototyping + Polis-like system”.
  • What is democratic process anyway? Definition by [OpenAI
    • What is a “democratic process?”
      • (A) a broadly representative group of people exchanging views and careful deliberation; and (B) a broadly representative group of people exchanging views and careful deliberation,
      • (B) Through a transparent [decision-making process
      • (c) It means the process that ultimately determines the outcome.
  • Voting without good debate is not a democratic process.
    • Need a place for careful deliberation with people from different clusters
    • A common “basis for discussion” is necessary for deliberative discussion.
    • How to make it? That’s where sci-fi prototyping comes in…
  • In software development, prototyping gives a more concrete picture of what needs to be improved.
    • But prototyping with software is something only software engineers can participate in.
    • So we need prototyping to create a story of the future in Japanese, this is SF prototyping.
    • With LLM’s support, even ordinary people who are not science fiction writers get the means to prototype the future.
  • The stories created form a “shared perception” that “this kind of future could happen.
    • A number of stories make Future Width visible.
    • This is the basis for discussing how we can make a better future.
    • image

Polis-like system

.

  • Polis is mentioned by OpenAI itself in the grant commentary, and as an example, a Polis-inspired system is presented, which is an important elemental technology in this field.
    • There was a preprint from the joint team of Anthropic and Polis the other day that you might want to read.
  • Polis is a system for visualization and clustering of “voting matrices” created by multiple people answering multiple questions with Yes/No answers.
  • image
  • Multiple people answer multiple questions to obtain a “distribution of people’s subjectivity” - From the subjectivity of one to the subjectivity of many
  • Our team is now using Polis as a simple user, but that makes it hard to try new things technically, so we are creating a new system: Japanese cockchafer (Melolontha japonica).

Repeat SF prototyping and Polis-like system

.

  • image
  • Gradual improvement through repeated iterations iterative structure.
  • Gathering and visualizing reader feedback on stories generates new ideas and creates new stories
  • All of these things accumulated and as a result…

Japanese cultural AI can be done.

  • AI can be created using a dataset of thoughts and emotions of Japanese speakers.
  • Many people may associate the term “Japanese culture” with classical culture, but the first thing I want you to imagine is…
    • Doraemon
      • A science fiction animated series that has been running for 44 years.
      • No one in Japan has ever heard of Doraemon, and the recognition of Doraemon is spreading quite well around the world.
    • Doraemon is creating a “AI is your friend” worldview.
      • The antithesis of the Hollywood-style “AI trying to destroy humanity” story
      • Those who envision a bad future based on the “AI trying to destroy humanity” storyline try to create legally binding regulations on AI.
  • Another keyword that came up from the project participants: the
    • mourning god
      • A human-made tool, after long use, becomes a soul and eventually becomes god.
      • Future image of LLM becoming wise and personable.
      • People in monotheistic cultures recoil at the idea of calling it “God.”
      • The Japanese concept of “God” is quite different from that of monotheistic religions.
      • Japan has only 1% Christianity, while the U.S. has almost 70%.
    • AI is a tool, an argument that should be regulated to prevent it from acquiring capabilities comparable to humans.
      • God made man.
      • God is a higher being than man.
      • Man created AI.
      • You think AI could rival humans?
        • “The created can line up with the Creator” would be a challenge to the concept of “God is a higher being than man,” irreverence!
        • Man has done God’s work of “making man.”
          • It is ethically undesirable, as is the cloning of human beings!
      • To make Global Deliberations possible, we need to understand each other’s cultural differences.
      • Action is needed to help the people of the world understand the concept of doraemon and attached mourning gods.

Technical feasibility of Japanese cultural AI

  • It may seem like a pipe dream, but it is highly technically feasible.
  • First, they use Scrapbox as a “place for communication”.
    • Anything can be written here, including generated stories, impressions, and other ramblings.
  • Realized implementations
    • Export data from Scrapbox once a day with Github Actions
    • Divided into chunks of 500 tokens and vector embedded with OpenAI Embedding API and stored in Qdrant
    • Web service that enables vector search with Qdrant is already available: Omoikane Vector Search.
    • Now we just need to create a connection between the English UI and the ChatGPT API, and we will have a Minimum Viable Product that can be released to the world.
  • Interesting points unlike a mere natural language corpus
    • image
    • From the Polis-like system we can get the data “37% agree with this question, 37% disagree, 25% abstain”.
      • Data on “the subjective distribution of a large number of people.”
      • Since natural language sentences are paired with quantitative data, by hooking the first half of the sentence with a vector search and loading the quantitative data into the prompt, the chatbot can speak with knowledge of the distribution of Japanese thoughts and emotions.
    • In the future, it would be interesting to build a mechanism to collect human responses by asking questions to a Polis-like system with Function Calling.

now in open beta!!!

This page is auto-translated from [/nishio/OpenAIの「Democratic inputs to AI」に応募した話](https://scrapbox.io/nishio/OpenAIの「Democratic inputs to AI」に応募した話) using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.

Graph View

Backlinks