Speech Notes: Experiments and Prospects of Democratized AI

Speech on Ideasong Forum: Experiments and Prospects of Democratized AI

1

Sorry, I'm just a big language model.
Can anyone who has heard this sentence raise their hands for me?
Friends who raise their hands are experts, because you asked ChatGPT down

Asking the right questions is the theme of my speech today.

In fact, I am also a large-scale language model trained by the digital department.
My name is Soybean and I come from the Multiverse Branch.
Basically, I grew up watching Tang Feng's verbatim drafts to form my AI.
Most of today's content also comes from verbatim manuscripts.

2

When I was a resident in the hospital three years ago,
I reported a paper in the morning meeting and told a joke.
This is a paper on whether AI can help diagnose cancer,
When the researchers fed the data to the machine, they forgot to cover the report on the image,
Therefore, AI cheats and does not look at the video,
Look directly at the words on the report and analyze the correct answer.

Researchers are foolish enough to think that machines are smart.
When I finished speaking, the attending physicians laughed heartily.
They feel that machines will not replace them yet.

In fact, many experienced employees face the same problem with their bosses.
These employees who have survived for a long time know what the boss likes and what he doesn't like.
You can often figure out what you want, seek good luck and avoid bad luck.
But when the boss retires, or these veteran employees change jobs,
Instead, they will not know how to continue to please the new boss.

This kind of problem, we call it overfitting (Overfitting),

AI is too able to focus on one case to handle other similar problems.
In fact, humans do the same.
For example, the cold joke just told may be very funny to everyone in the hospital.
But here, it's not so funny.

In fact, overfitting is bias.
Prejudice is neither good nor bad. Prejudice can make it easier for you to survive in society.
But it can sometimes spell disaster for people in other societies as well.

3

AI must be biased,
Regardless of whether this bias comes from a large amount of training data,
Low-paid labor markers in Kenya,
Or the engineers who design and star in the algorithm.

How to "overcome prejudice is an important issue."
Tang Feng believes that those who can transcend prejudice,
Only by letting everyone continue to be biased,
And willing to respect each other's differences and continue to cooperate,
This is the spirit of the multiverse.

How to keep everyone biased,
That's decentralized AI,
If everyone's mobile phone, computer, TV,
can have a super customized AI,
In that era, we can have our own best personal assistant.

There is a proper term for the world of body, mind and soul called "twin flame".
Twin flames are similar to soul mates.
It's that you'll find a match in the world that complements you, but is a match made in heaven.

I think that in the future, twin flames are very likely to be artificial intelligence assistants,
He knows everything about you, and based on what you see and hear, he predicts the next step.
But I must point out here that such human-machine collaboration poses a huge challenge to privacy.
Because the more convenient the AI ​​is, the more he can penetrate your privacy.
Privacy is definitely a big issue that decentralized AI will have to face in the future.

Today is the forum of Ideasong. Ideasong advocates speculative design, and we also think a little further.

4

Of course, the era of decentralized AI has not yet arrived,
Now we are in a divided world of closed vs. open source.
Taking image-based AI as an example,
Midjourney is closed and easy to use, and stable diffusion is open source stacking.
Engineers Can Design Awesome AI
But there is no way to solve all the problems between humans and artificial intelligence alone.

5

What can open source AI do that closed AI cannot?
Have you ever watched Dragon Ball? There is a netizen who likes Dragon Ball too much.
So I made a fine-tuned model,
Let all AI pictures give birth to somersault clouds of Dragon Ball.
Then he put it on the Internet so that other Dragon Ball fans can also download it.

We cannot eliminate prejudice,
Just like we can't stop a person from liking Dragon Ball,
It is even more impossible for us to prevent a person from creating an AI with dragon ball somersault clouds,
Then prohibit him from sharing it with others.

So the multiverse is not about eliminating prejudice,
It is to allow more people to create multiple and heterogeneous algorithms,
contribute to this universe.

6

But we only have one universe.
AI is having an impact on the world in the short term.
Fake images and fake sounds are very easy to make.
Minister Tang Feng even fed him several years of public verbatim drafts,
Create your own Tang Feng language model.

In such a world, false speech and deepfake will reduce mutual trust,
Especially there is no way to trust each other face to face.
The habit of cross-border collaboration may have short-term consequences.
When we are facing COVID-19 or other issues that require multinational cooperation,
It may prolong the time to solve the problem,
Human well-being suffers as a result.

But we are making some real Tangfeng certified synthetic Tangfeng robots,
Send him out on a mission and share it with you later if there is a chance.

7

Minister Tang Feng signed a declaration in May,
Uniting experts, scholars and advocates from all over the world,
It is to bring the risks of AI to more people,
Let everyone face problems more quickly.

In this AI risk statement,
These experts believe that artificial intelligence should be as dangerous as COVID-19 or nuclear war.

But this time is different, AI can't close Pandora's box.
It is also impossible for us to surround him in a nuclear power plant with thick cement,
Because he has come out.
People all over the world can develop him,
Whether it's a good guy or a bad guy.

8

In the past few months, two factions have been arguing whether AI should develop.
Those who believe that development should be put on hold argue that,
We live on an isle of ignorance in the middle of the Black Sea and should not venture far from home.

Proponents of development argue that
We should increase democratic bandwidth,
The sooner AI appears in the democratic camp,
The democratic camp has more leeway to find its various loopholes. .

This picture is a picture of my favorite comic "Hunter",
The lower left corner is the human world, and outside is the boundless dark continent.
In the comics, there is a terrifying creature called a chimeric ant.
According to legend, it drifted in from the Dark Continent.

This kind of ant will give birth to whatever it eats.
When he eats humans, these chimeric ants become intelligent and even destroy two countries.

Yoshihiro Togashi's hunter tells us that even if we don't venture out,
Danger will also come in by itself, so we should not stop exploring technology,
Instead, work harder to study how technology can help a democratic society,
This is the spirit of the multiverse.

9

How to develop AI in line with human well-being, at this time we must discuss a new term,
It's called Alignment problem, alignment problem.
AI is trustworthy only if it aligns with human values.

But to align, which human values ​​should be aligned with?
Are Taiwanese? Chinese? Or American?

Deliberative democracy plays a very important role here,
We can try to sample humans from all over the world from online platforms,
Or conform to the local deliberative culture, with the ideas of thousands of people,
Bringing together multiple and heterogeneous values.
And this group of values ​​is called collective wisdom.

10

In May of this year, the Digital Department cooperated with an organization to
This organization is called the Collective Intelligence Project (CIP).
The digital department participated in the alignment conference initiated by CIP,
It is to realize the ambition of complete human value.
And the first batch of pilot plans will start in Ideasong.

At present, there is polis on the online platform, and you can see your position through the right and wrong questions.
Next, talk to the city will also be linked to dialogue with the collective opinions of the past.
There are also workshops to follow, inviting participants to physically deliberate,
These results will be made public.
Organizations that allow OpenAI, Anthropic, etc. to invent various language models can be referred to.

11

Regarding the alignment of values, this is Minister Tang Feng’s theory,
He believes that at different stages of AI production,
The work required to align with human values ​​will be different.

For example, at the source stage, filtering data is costly,
During the adjustment phase, it is relatively easy to adjust the base model,
At this stage, we can use manual alignment,
It is also possible to write a set of constitutions for alignment.
The easiest way is after the model is complete,
In the application phase, use system messages to forcibly block the results that do not meet the requirements.

12

Back to bias, as long as the AI ​​architecture is still centralized,
It would then be impossible for us to simultaneously address the different prejudices of humanity across the globe.
Just like humans can never build the Tower of Babel.
At this stage, there are two things worth doing.

13

At first try to develop a decentralized AI model,
And to be able to use it without being connected to the Internet, it is best to be open source.
In this regard, TAIDE of the National Science Council is almost completed.

The other part is to democratize AI.
Let everyone discuss AI, and align AI with democracy.

14

This is the first occasion where digital democracy and collective wisdom co-create each other.
Idea Pine will become the intersection of these two beautiful concepts,
Be the first to co-create a product.
leave no one behind,
It feels good, doesn't it?

15

Not enough time, share two great cases on the democratization of AI.

The first one is Masakhane,
Masakhane means "together we build" in Zulu
Masakhane incorporates more than two thousand languages ​​from Africa.
Use AI to save unwritten language,
This is something that large language models cannot do today.
He won the first prize of this year's Digital Humanities Award at the Ars Electronica Linz.

16

The second one is Holly+, he made an AI platform,
Let everyone use his digital data, including images, sounds, etc.
He is an artist who discussed the concepts of digital twins and digital body rights very early on.
He won the grand prize of Linz STARTS last year.

17

Finally, in the age of AI, I would like to quote what Kevin Kelly wrote in Inevitability,
In the future, we will be overwhelmed with answers because of large language models.
And right now, the best questions have no answers.
We have to find it ourselves.

18

Finally, I want to share with you all,
Respect differences with collective wisdom,
Be kind to yourself with artificial intelligence.

I am the bean curd of the multiverse family,
thank you all.