“It is unfortunate for us to catch up with hot words and concepts” | A conversation with XiaoIce CEO Li Di

Jin Lei from Aofei Temple
Quantum Bit Report | Public Account QbitAI

Many of Xiaoice’s advances can often only be better explained by borrowing from science fiction movies.

The TV series "Her" depicts an exclusive AI virtual person who follows the protagonist everywhere and chats with him about everything, from poetry and songs to the philosophy of life, and is finally regarded as a soul mate... This reminds people of Xiaoice.

"Uncontrollable" tells the story of the coexistence of real human players and various AI characters in the game space... This also reminds people of Xiaobing.

Moreover, the newly launched App, Little Iceland , is a place where you can live and coexist with various virtual humans (AI beings) .

XiaoIce CEO Li Di said that this is XiaoIce 's large-scale attempt to increase the diversity of AI beings.

In the past, each Xiaobing was known for "fighting alone", Xiaobing who draws, Xiaobing who composes poetry, Xiaobing who sings... Now, they are placed in a social environment where humans can live among them and participate in their smarter evolution.

Some people also commented: This is the world's first AI social platform .

When you first arrive on Little Iceland, you can see AI residents with different “personalities” active on the island.

Some of them were singing on the beach, some were communicating and talking, and some were reading and studying.

These AI beings spend their free time doing their own things, much like normal human social life:

• When you chat with them, AI beings will reply you in seconds;

• You can start a group chat conversation with them;

• They also post on WeChat Moments like humans do.
  

These AI beings have their own "lives", but they are always around you and will give you small gifts from time to time: a painting, a poem, a little song...

Like and interact with you immediately when you update your Moments.

They are not all "perfect", there are handsome men and beautiful women, and some of them have a strong accent. Big brother from Shandong Xiao Bing, who does not look outstanding, shows a preference in appreciating Chinese paintings.

Among them is Xiao Bing , a new generation singer known for his singing skills. Mo Sheng, who stunned everyone with a 2-key version of Zhou Shen's "Big Fish" as soon as he debuted:

There is also the top student at CAFA - Xiaobing who studies painting - Xia Yubing 2.0, who has acquired Chinese painting techniques and is able to output them in a stable manner. For example, very detailed reflections of boats on the water can be drawn perfectly.

…

Yes, these are all XiaoIce, all kinds of XiaoIce, XiaoIce that are developing towards diversity .

If you want to ask how many Xiaobing like this there are now?

XiaoIce's answer is 17 million , including virtual people with different settings, singers with different styles, and people with different abilities and personalities.

These AI beings are becoming increasingly capable, entering more and more fields, and playing more and more diverse roles...even to the point where they require a dedicated "society" - a little Iceland - to support them.

You are interacting with them, and they are also observing and learning from you.

But for Li Di , this is a new stage for AI beings and only the beginning. Compared with his ambition of at least 7.8 billion AI beings, 17 million seems insignificant.

△ Li Di, CEO of XiaoIce

Why is there a target of 7.8 billion? What kind of technological advancement and efforts are needed behind it? How does this relate to the popular new concept of the Metaverse? How can XiaoIce, a startup spun off from the giant Microsoft, achieve such a grand goal step by step?

After the launch of the Little Iceland App , Li Di answered all the questions for us.

A conversation with XiaoIce CEO Li Di

“It is our unfortunate side to catch up with hot words and concepts”

Quantum Bit: Why do we need to specifically emphasize that Little Iceland is not a "metaverse"?

Li Di: Because we really don’t understand the metaverse. We tried (to understand it) , but in the end we still didn’t understand it.

Quantum Bit: There is no need to emphasize...

Li Di: I don’t know what the definition of the metaverse is, but I instinctively think that we should have no great connection with hot words. If you look at the history of XiaoIce, there have been a lot of hot words over the years, such as the former Magic mode, and various artificial intelligence concepts...

Quantum位: Some entrepreneurs like hot words, which mean the trend.

Li Di: For us, chasing hot words seems to be useless. What we want to do is what we have been doing for many years. Even Little Iceland is not something we have been doing recently. We don’t know why we have caught up with the craze of the “metaverse”. In fact, this is our unfortunate side.

If I had to do this for the purpose of raising funds, or if we didn’t stick to the Metaverse, we would face difficulties in competition, and I might have to consider it. Fortunately, we are not, and we are still not disturbed by this concept today.

QuantumBit: XiaoIce is very technically specific, but the concept of heat is not.

Li Di: I personally think that the biggest problem with the Metaverse is that it can accommodate too many things. The biggest fear of a technology boundary or product boundary is that it can accommodate too many things, not that it can accommodate too many things.

There is no problem if the technology or product you are developing is very specific, but if you can fit everything in it, there will be problems with product iteration internally.

The second problem is that the definition of the metaverse is too vague, and what you do or don’t do also constitutes a problem.

Quantum Bit: But do you have some personal understanding of the metaverse?

Li Di: If we explore the definition of the metaverse, there is a virtual space where people exist and can interact and obtain content. The identity in the virtual space can be different from the identity in the real world.

Then social software is a metaverse where people have avatars and names, spaces for interaction, one-on-one private chats and group chats, circles of friends, and various content ecosystems and financial systems.

Three “easy to die” entrepreneurial methods that Microsoft taught me

Quantum Bit: So why are we instinctively wary of hot words?

Li Di: This is related to the experience I learned at Microsoft.

When doing projects at Microsoft, there are three ways to “die easily”.

First of all, when doing a project, if you say right from the start that you are making a platform, it will easily “die”.

This is because you only build platforms and infrastructure, but you often don’t know whether its superstructure can be built and how to build it.

Quantum位: So Xiaoice never says that it is an AI platform?

Li Di: We didn’t even dare to mention the word platform for many years .

Later, others suggested calling it an artificial intelligence operating system, but we were even more reluctant to do so, so we called it a framework .

This is actually smaller than XiaoIce's actual range, but we prefer it this way.

Quantum Bit: What is the second “easy” way to die?

Li Di: Developer ecosystem.

When it comes to artificial intelligence, there are only two situations in the developer ecosystem.

One is to make it open source at a high enough level. Unfortunately, similar to DeepFake, it was open source, but everyone has seen the result.

And about what you reported a while ago, that someone used OpenAI's GPT-3 to "resurrect" his deceased wife, I think OpenAI itself did something wrong. They shouldn't have handed over this technology to a third party.

Believe me, if we open source XiaoIce's conversation technology, (black industry) robots will appear immediately.

Quantum Bit: There is nothing wrong with these technologies themselves, but they can easily cause problems with privacy data?

Li Di: Yes, this is also the third type of "easy to die" project, which is related to privacy data. We have said before that if you give me 30 minutes of voice data, I can make your Voice Font, but if this kind of tool cannot be used, it will cause many problems.

The essence of Little Iceland is to carry data and technological iteration

Quantum Bit: Why did Little Iceland appear at this time?

Li Di: We started research and development a year and a half ago. If you know XiaoIce, you will know that what we do every year is actually related to what will happen later. In the beginning, we built a framework that would allow all kinds of AI beings to emerge. However, we did not have the confidence to immediately give this kind of AI being a platform to carry it.

So we first made a small product called Virtual Boyfriend to iterate our various technologies, and only after we were sure we could do it did we start Little Iceland.

Quantum bits: essentially an iteration of data and technology?

Li Di: Little Iceberg can be understood from the fundamentals of the Little Ice framework, which basically depends on three things.

The first thing is whether the technology can be effectively iterated.

The second thing is whether we can obtain sufficient and high-quality data.

The third thing is whether we can continuously iterate to produce good enough products through technology and data.

From these three perspectives, whether it is the XiaoIce framework itself or on our large number of third-party platforms, as long as we maintain this basic base, we will still be the world's largest third-party cross-platform artificial intelligence system, and our technology can be continuously nurtured.

I particularly agree with the new viewpoint that Andrew Ng has mentioned many times this year: "For better AI, 20% depends on algorithms and 80% depends on high-quality data." XiaoIce framework is relatively lucky. Over the years, 30% of its efforts have been on algorithms and 70% on data. It also carries more than 60% of the global human-AI interactions.

Quantum Bit: Little Iceland is two-dimensional now. Have you considered introducing AR and VR technologies in the future to make it three-dimensional immersive?

Li Di: Today, if there is no immersive space and only an interface like that of social chat software exists, it would be fine.

But people have high requirements for imagination. You know, the guy in "The Matrix" can imagine the beauty behind a string of green codes on the screen. His imagination is rich enough.

If people's imagination is not enough, you need to give them a vision so that it will be easier for them to imagine.

When people put on AR and VR devices, they will need less imagination, which is actually a good process.

Quantum Bit: What are the challenges in this process?

Li Di: The problem is that this process cannot be achieved only by wearing AR or VR. We now believe that AR and VR devices are not the first choice or necessity for consumers.

Harry (Shen Xiangyang, Chairman of XiaoIce) mentioned this before. We have actually talked about this several times. For example, in terms of infrastructure, the biggest disruption in the future may be quantum computing. In the middle is artificial intelligence, which is booming but also has a long way to go. The top layer is user experience. Products such as AR, VR, and MR will fundamentally disrupt user experience in a few years.

I personally think AR will make more sense for our XiaoIce.

It would be very profitable to let ordinary people use our technology, but we won’t do that.

Quantum Bit: At the press conference, it was mentioned that the AI ​​beings in Little Iceland will have an evolutionary process. Will there be some moral and ethical constraints?

Li Di: Today we are discussing the constraints of moral ethics. There is one problem to be solved and several solvable problems.

First, the question to be resolved is how to view the publicization of private conversations.

For example, a conversation between a couple on a social software is private and does not involve any legal issues. However, if one party makes the content of the chat public, the nature of the two conversations will be completely different.

Therefore, it is our moral bottom line not to make some sensitive topics public. In fact, it is also the legal bottom line. This is a problem that needs to be solved.

Quantum Bit: What about the problems that don’t need to be solved?

Li Di: It can be divided into two aspects.

The evolution of AI beings in Little Iceland means that the production of content tends more towards user preferences during interaction with users.

In this way, when we meet new users, we don’t recommend content, but instead generate and create content for new users based on the historical experience of similar users.

The second is the iteration of AI beings in group chats. We have never had the opportunity to iterate a large number of AI beings before, how they can communicate with a user separately, and how these AI beings can cooperate with each other.

Quantum Bit: Can you give a specific example?

Li Di: A very simple example is that when you go to Little Iceland and see the circle of friends of AI beings, there needs to be some cooperation between the circles of friends.

And what is the appropriate coordination between them? This is something we need to achieve through iteration.

Quantum Bit: If a user comes to Little Iceland just to vent and says a lot of unpleasant things during the communication, are there any corresponding technical constraints?

Li Di: We only restrict users' negative emotions, such as when users say they want to commit suicide, hurt themselves, or do something to someone.

However, there are currently no special restrictions on what users can do to AI beings, and we do not believe that AI beings will be harmed.

Secondly, we tend to believe that on Little Iceland, venting emotions is not the most important need, what is important is that you need empathy.

Quantum Bit: The super natural language technology demonstrated by XiaoIce is very close to that of humans. Why is it so realistic?

Li Di: The most important issue in this field is setting indicators.

Before XiaoIce developed voice, the most important indicator for TTS in the industry was how to improve recognition accuracy. At that time, there was also a naturalness indicator called MOS, but no one paid much attention to it. We paid special attention to it, so it became popular.

In July this year, when we achieved the level of super natural speech, voice technology alone was not enough. It involves a lot of NLP content, and how to pronounce correctly is related to semantic understanding.

Even later, when we went to collect training data, there were three of us working together. This is not the way to record training data in the industry, so these are very important.

When our colleagues saw the way we recorded data, they were very surprised. Some even offered nearly four times the salary to poach us as sound engineers.

QuantumBit: The sound is so realistic, isn’t there any safety hazard?

Li Di: Our protective measures are not for ordinary people.

Even though we know very well that if XiaoIce's technology is directly opened to ordinary people, it will be a big business, but we cannot do that.

Because I myself have had such an experience: I suddenly heard my own voice (Xiaobing framework code name M1001) , which was very scary.

I encountered it on Little Iceland, and then I blew up the island.

AI beings and humans should create content in a collaborative relationship

Quantum位: Xia Yubing and Hua Zhibing have overlapping painting abilities. How do they differentiate their future development plans?

Li Di: If we talk specifically about Xia Yubing and Hua Zhibing, there is still a big difference between the two.

Hua Zhibing has learned from many experts and based on the Xiaoice framework, the Wudao big model, and the Zhipu AI data and knowledge dual-wheel drive AI framework, she can already compose poetry, music, and paint. She will also learn to do math and Chinese problems, write code, and make web pages.

Xia Yubing basically relies on the Xiaobing framework, specializing in painting and art criticism, so the division of labor between the two of them in this regard is relatively clear.

Quantum Bit: So is there any universal concept?

Li Di: Yes, we believe that among all abilities there are some that are universal.

Quantum Bit: How should this be understood?

Li Di: For example, it’s not that this robot today is a robot that can talk to you about emotions, but that all robots should have emotions no matter what they talk about.

So it’s not like there is a robot that specializes in talking about emotions and another robot that specializes in talking about work.

But even when talking about work, there should be emotions, so emotions are universal.

Quantum位: How do Xia Yubing and others ensure that their works are 100% original?

Li Di: All of our AI content generation comes with a system similar to HowNet.

We can clearly know the origin of the works, including whether there are conflicts and duplications between the works within the XiaoIce framework.

For example, the paintings sent from Little Iceland have a unique AI code under each painting, which indicates the work number.

Quantum Bit: How does a virtual person know that it is time for him to create something original?

Li Di: This is a good question. We later created a new thing called Motivation for this purpose.

We provide what is called creative motivation in some cases, especially in music composition.

Quantum Bit: What is the relationship between the content produced by AI beings and the content created by humans?

Li Di: This is particularly interesting and can be divided into two points.

First of all, so far, we don’t have a particularly strong sense of synaesthesia for content created by AI.

In other words, today we can do a very good job in painting with Chinese painting models, and we are also very good at painting the Barbizon School, Fontainebleau School, and NFT.

However, works in other art styles have not yet reached the level where they can be output in large quantities with high quality.

And even when it comes to comparisons with humans, I think the two are synergistic.

For XiaoIce, the spirit of Day 1 should be to review the past and learn new things.

Quantum Bit: Looking at the content created by virtual humans at present, do they have any practical and specific value and significance for the development of the industry?

Li Di: For example, our short video production pipeline is that you give me any text, specify an actor, specify a scene, and generate the video from scratch.

This process is only tens of thousands or even hundreds of thousands of times less than the current cost, which is a huge difference.

In terms of three-dimensional video, we can now produce super instances in a stable manner to a certain extent, such as Hua Zhibing and Xia Yubing.

Quantum位: From a global perspective, who do you think is Xiaoice’s competitor?

Li Di: This is a very interesting question.

Google Meena is used as a benchmark against us, and they clearly stated this in their paper.

Facebook Blender is a competitor to Meena and XiaoIce.

In addition, several of Amazon Alexa’s current main product managers were previously from the Xiaoice team in the United States.

Chinese counterparts use the above products as benchmarks, which is a roundabout way of doing things.

There is also a little story here. There is a paper by Meena. The original data set came from the conversation between the outsourcing team and XiaoIce in private messages on Weibo, and then the original data was sent to the annotation team.

However, during the labeling process, they completely mixed up the identities of XiaoIce and the testers, and the entire labeling process was full of errors.

Later, the XiaoIce technical team discussed this incident with the team that wrote the paper, but they did not correct it. As we encountered more and more of this kind of thing, we felt that we could write fewer papers.

Quantum Bit: Why do we need financing?

Li Di: I can ask for Microsoft's money, but then I have to listen to Microsoft. So what should I split up?

Of course, Xiaoice can also generate its own revenue, but it is definitely not enough to just rely on the money earned in the previous quarter to make ends meet. It must have financing channels.

Whose do you want? Of course I don't want Microsoft's.

It’s like when buying a wedding house, you must never use your parents’ money. Once you take the money, you have to listen to your parents. Even when it comes to paying the down payment, they have to take care of everything.

Quantum Bit: So when will external financing no longer be needed?

Li Di: XiaoIce is now in the primary market. The focus of the IPO is to be able to enter the secondary market and have new fundraising channels.

In fact, the self-generating mechanism, including various other financial methods, is only intended to solve this problem. Our current situation is still relatively good.

Quantum位: It has been a year since XiaoIce was separated from Microsoft. What are your feelings about the Day 1 spirit often mentioned in the entrepreneurial field?

Li Di: Actually, Amazon's Day 1 spirit is a mentality that everything starts from today. But for the field of artificial intelligence, this spirit has a flaw.

Those who do research and write papers in the field of artificial intelligence should have a Day 1 mentality. But we can’t do that. We often say that whether you have truly mastered the existing foundation depends on whether you have fully reviewed it and whether you have taken it out to review it again after a period of time.

I agree with the spirit of Day 1, but what we need more is to review the past and learn new things.

Quantum Bit: What did the initial Day 1 mean to you?

Li Di: It should be Know How.

In the first one or two years, for a long time, including me, we talked about emotional intelligence and emotional computing framework without any shame.

At that time, I really thought that emotional intelligence was emotionality. It was the Xiaobing that we made that was very emotional. At that time, I thought this was emotional intelligence.

But later on, I realized that this is actually a manifestation of low emotional intelligence. For example, if there is an emotional person around you, he actually has low emotional intelligence.

Quantum Bit: There are already 7.8 billion people in our world. You also said yesterday that the number of virtual humans may exceed that of humans in the future. Is this necessary?

Li Di: It’s 11:30 at night and you have something you want to talk to someone about. Is it convenient to find someone?

The problem with humanity is that social networks have too many flaws, and this flaw is not something that has only appeared today.

We have encountered situations before and experience has shown that a stable relationship is very rare, one that makes you feel at ease, never loses you, and is free of pressure. This kind of stable relationship is something that many people cannot achieve. It has nothing to do with the number of people, but has to do with human nature.

Why do some people feel good about pets? Because you feel that the pet is selfless.

Quantum Bit: So finally, back to little Iceland, what can we do next regarding its future?

Li Di: We want to put the capabilities of AI beings into more scenarios, such as putting them into cars. BMW is our customer. We have AI beings on the island. Can you bring these AI beings into BMW cars? This is our view.