Zhang Yijia, CEO of Lingxin Intelligence: There is no need to do China's OpenAI!

Focus:

01

Factual errors and other problems in large language models can be solved in the long run, either through engineering methods or directly through models, but this will take a certain amount of time and will be difficult.

02

At present, it is the closest moment to general artificial intelligence, but large models are not necessarily the only path to general artificial intelligence, and the future may be implemented around large models and other combinations.

03

It is difficult for a large base model to meet all needs, both the base model and a certain domain model, through the combination of ecology, better to form a closed loop with the application and user needs.

04

Domestic companies do not need to benchmark China's OpenAI, and the opportunities for startups lie more in the direction of clear needs rather than consensus.

Introduction of the panelists:

Yijia Zhang, CEO of Lingxin Intelligence, holds a bachelor's degree in computer science from Tsinghua University, a master's degree in computer science from Carnegie Mellon, and a double bachelor's degree in economics from Peking University. The founding team of Google voice assistant, the founding team of Android Auto, used to work at Google, participating in search, advertising, payment, car, email, enterprise project management tools and other projects.

Produced | Sohu Technology

Author | Liang Changjun

In the wave of domestic large-model entrepreneurship, many enterprises carry the imprint of universities, among which Tsinghua can be said to be in the first echelon.

Lingxin Intelligence, which completed two rounds of financing in less than three months, was incubated from Tsinghua's computer science department, founded by long-term associate professor Huang Minlie, and the company's CEO Zhang Yijia, also graduated from the department.

In December last year, Lingxin launched its first product, AI Utopia, through which users can talk to custom AI characters. Behind this is the super-anthropomorphic large-scale pre-trained language model self-developed by Lingxin Intelligence. Zhang Yijia recently told Sohu Technology when attending the 2023 Sohu Technology Summit that the capabilities of this model have largely reached the level of ordinary people, and even better in some scenarios.

For the factual errors and other problems of large language models, Zhang Yijia said that these problems can be solved in the long run, but from the perspective of anthropomorphic scenarios, it is not absolutely required that the model is completely correct. "Sometimes its uncertainty, on the contrary, brings a kind of entertainment, which creates the feasibility of applying many scenarios."

He said that Lingxin Intelligent targets foreign Character.AI, and is not inferior in the same Chinese content. Founded in 2021, Character.AI allows users to create custom AI companions for entertainment, role-playing or emotional support, with a latest valuation of $1 billion.

There have been many imaginations about AI companions. In the sci-fi romance movie "Her", the girl named "Samantha" - an AI operating system, humorous and understanding, eventually developed a love relationship with the male protagonist, but in the end, the departure of "Samantha" made the male protagonist deeply sad.

In the future of AI, whether AI will have conscious emotions and the relationship between people and AI have been widely discussed.

Zhang Yijia said that Lingxin Intelligence has added a lot of empathy, companionship and other corpus to model training, which can be expressed, but whether it can be understood from the logical level or psychological level, to achieve the emotions that people have, is actually difficult to explain. He believes that AI still needs to do some work to reach the state of AI in Her.

However, Zhang Yijia believes that it is indeed the closest moment to general artificial intelligence. "But whether the big model is the only path, I don't think it's necessarily." He believes that in the future, general artificial intelligence may be implemented around a combination of large models and other methods.

To some extent, it is precisely because the big model is regarded as a key to opening general artificial intelligence that it has attracted domestic and foreign technology giants and a number of startups including Lingxin Intelligence.

Zhang Yijia said that at present, there is no large model of the base to achieve general artificial intelligence, and it is difficult to meet all the demands with only one. He believes that both the base model and a certain domain model are needed, and eventually it may be through the combination of ecology to better form a closed loop with applications and user needs.

"But not all of them have to benchmark China's OpenAI, I think it's actually wrong." If you do this, at least from the perspective of the startup team, it is difficult to fight too many big factories, as well as OpenAI. In this big model boom, Zhang Yijia believes that the opportunities for startups lie more in the direction of clear needs rather than consensus.

He also said that he hopes to co-create with users to conceive the future relationship and communication mode between AI and people, so as to gradually see the possible development direction between people and AI in the process of exploration.

The following is the transcript of the conversation between Sohu Technology and Zhang Yijia:

Sohu Technology: You mentioned the super-anthropomorphic large-scale pre-training language model, what is its current anthropomorphic level? Already surpassed people?

Zhang Yijia: It has surpassed people in some scenes, for example, I am more steel straight man, in simulating the role of warm man, it is actually more able to chat with girls than me. It has largely reached the level of ordinary people, or even better than ordinary people.

Sohu Technology: From this point of view, how big is the gap between China and abroad?

Zhang Yijia: In the matter of anthropomorphism, our work is not weaker than that of foreign countries. If we talk purely about functionality, we have a certain gap, but in terms of anthropomorphism, we are targeting foreign Character.AI. On the same Chinese content, we are not inferior to it.

Sohu Technology: Large models realize automatic generation of copywriting, code, etc., but logical thinking is sometimes poor, or factual errors occur, how to solve this aspect?

Zhang Yijia: In the long run, this can definitely be solved. Some search companies take an engineering approach to solving the illusion problem generated by the emergence of a model, which is equivalent to making a model based on the generated content to monitor whether it is correct. A more beautiful approach would definitely be to solve it directly through the model, which may take a little time and will be difficult.

From the perspective of anthropomorphic scenarios, in fact, it is not absolutely required that the model is completely correct, after all, people also make mistakes. In the anthropomorphic scene, this is not the core problem, sometimes its uncertainty, but brings a kind of entertainment, which creates the feasibility of many scenes to apply.

Sohu Technology: In the long run, will large-scale pre-training language models finally have human consciousness or emotions?

Zhang Yijia: From the current point of view, it will have memories, and then it will also express the ability to empathize in some occasions. In the model we do, a lot of empathy, companionship and other corpus are added to the training process, which can be reflected, but whether it can be understood from the logical level or psychological level, to achieve the emotions that people have, it is actually difficult to explain. To achieve the state in the movie "Her", it still takes some work to do.

Sohu Technology: Many views believe that large models are a path, or even the only path, to general artificial intelligence, and will large-scale pre-trained language models accelerate the arrival of general artificial intelligence?

Zhang Yijia: In the matter of general artificial intelligence, AI practitioners have experienced several waves, and this time it really feels the closest. But whether the big model is the only path, I don't think it's necessarily. I think that in the future, general artificial intelligence may be around the big model and the extra work on top of the big model, or it will be implemented in a combined way.

Sohu Technology: At present, domestic Internet giants, including many startups and scientific research institutions, are making large language models, how do you view this wave? Will there be so many big models needed in the future?

Zhang Yijia: If a large pedestal model is really good enough for everyone to use, customize, and meet the needs, then it doesn't have to be so much. But at present, it has not yet reached general artificial intelligence, and it all requires personalized needs and experiences, and it is difficult to say that a large model of the base can meet the demands of everyone. Therefore, you need a large model on the base, and you also need to combine the data in your own domain to train a model in a certain field, and eventually it may be through the combination of ecology to better form a closed loop with the needs of applications and users, I think this matter is more important.

Sohu Technology: What are the opportunities for startups in such a boom?

Zhang Yijia: Startups should look for some direction that sees demand, but is not non-consensus. If they all think of the needs of consensus, startups to do it, whether it is human, material, or financial, are actually not dominant, and to a certain extent, it is also a waste of resources.

But not all have to benchmark China's OpenAI, I think it's actually wrong. If you do this, at least from the perspective of the startup team, it will be difficult to fight the big manufacturers, as well as overseas OpenAI. There are many scenes that are not absolute demands for productivity, as well as people's emotional demands, customized companionship, etc., which may be a better entry point.

Sohu Technology: How is the positioning of Lingxin Intelligent different from other companies that make larger models? What are the technical advantages?

Zhang Yijia: We have been doing dialogue and emotions for a long time, how to represent emotions, express emotions, how to integrate emotions into the model, how to make the emotional intelligence of the model higher, behave more like people, so that it can be thousands of faces, so that everyone can shape the general artificial intelligence companionship they want, which is our advantage.

We hope that users will co-create with us to imagine what kind of relationship and communication mode between AI and people should be in the future, so that in the process of exploration, we can gradually see the possible development direction between people and AI.

Sohu Technology: The risk of AI or large models is also a matter of concern, for enterprises, how to promote the development of AI while avoiding or handling its risks, and achieve a balance between the two?

Zhang Yijia: Our core work is how to make it more controllable. In the controllable dimension, it involves not only verbal security, but also ethical security, that is, values. Around the security controllable, we made a security framework, did a security evaluation, based on such a framework control model, so that it ensures its security in all dimensions.