Organized by Mingmin Xiaoxiao from AIGC Summit
Qubit | Public Account QbitAI
AIGC is so hot that if you don't keep up, you will be eliminated.
This is the most intuitive feeling in the current venture capital circle.
But "keeping up" is not an easy task: how to keep up? In what direction?
If you are not careful, you may miss the tuyere.
For start-up companies, AIGC technology is developing too fast nowadays, even if starting a business from the application layer, the company's PPT may not survive an OpenAI technology iteration;
For investors, there are too many AI new technologies and new entrepreneurial directions that have emerged in a short period of time, and learning from 0 is likely to pass by potential stocks.
Under the carnival of technology, companies and investors have different thoughts on the differences between large-scale entrepreneurship at home and abroad, the future development direction of large-scale models, and the innovation of AIGC.
To this end, Qubit invited Zhu Lei, the co-founder and COO of Yuanyu Intelligence, Chen Shi, the investment partner of Fengrui Capital, Ma Qianli, the co-founder of Unbounded Ai, and Lin Laini, the vice president of commercialization of Huayuan Shuzhiren, to discuss together Under this wave of AIGC, China is facing new opportunities and new challenges.
The roundtable session was hosted by Jin Lei, editor-in-chief of Qubit. On the basis of not changing the original meaning, Qubit has edited the content. Hope to bring you more inspiration and thinking.
The China AIGC Industry Summit is an industry summit hosted by Qubit, and nearly 20 industry representatives participated in the discussion. There are 600+ offline audiences and nearly 3 million online audiences. It has been widely reported and concerned by dozens of media including CCTV2 and BTV.
topic points
AIGC has allowed many talents to return to the AI track, which will be very beneficial to my country's technological development in the long run.
This round of AIGC upsurge makes entrepreneurs feel anxious, but more of them are moving forward in anxiety.
Most of the opportunities for AIGC start-up companies may still be at the non-model layer, or do some vertical models.
As long as AI does not betray humans, it will definitely bring about a soaring increase in social efficiency and human experience.
Becoming a company like OpenAI does not mean creating a Chinese version of OpenAI.
The huge wave of AIGC in China is said to be the "iPhone" moment, but it is actually more like the early PC Internet moment.
The following is the full text of the forum conversation:
Forum record
AIGC boom brings shock, excitement and anxiety
Qubit Jin Lei : I would like to thank everyone for participating in the roundtable forum on "New Opportunities for AIGC in China". Today, ChatGPT has triggered a big wave of AIGC, especially in China, which has formed a prairie fire. Therefore, we believe that at the current point of time, it is necessary to summarize and discuss the rapid development of AIGC in order to deal with the upcoming new opportunities.
To this end, we invited several corporate guests who are at the forefront of the industry to discuss this topic together.
The first is Zhu Lei, the co-founder and COO of Yuanyu Intelligence . Yuanyu Intelligence is the forerunner of this wave of AIGC upsurge.
The second place is Chen Shi, investment partner of Fengrui Capital . Cutting-edge technology has always been the focus of Fengrui Capital's investment.
The third is Ma Qianli, the co-founder of Wujie Ai . In terms of generative painting, Wujie Ai is also a pioneer among domestic companies. Currently, the number of users on the APP has exceeded 2 million.
The last one is Lin Laini, the vice president of the commercialization of Huayuan AI . Huayuan Computing is deeply involved in the field of artificial intelligence and AI, and has long been committed to the intelligent technology empowerment industry.
In this discussion, we set a keyword around each topic. The first topic is relatively easy, and the key word is " feeling ".
We would like to ask everyone to talk about this wave of AIGC upsurge, what is the most intuitive feeling brought to your industry and related businesses?
Yuanyu Smart Zhu Lei : The most direct feeling is shock and impact.
Although we have been doing large-scale model exploration and development in the industry, there will be certain expectations, but there are still many aspects in this round of upsurge that exceed our predictions.
One obvious aspect is the acceptance and application depth of ChatGPT in various industries. This is a very big change since the development of artificial intelligence. It turns out that AI is applied in small circles or fixed links. This time, it has truly broken the circle. From breadth to depth, it is a very big technological change. This is what we think is shocking.
For our industry, there are probably two impacts.
One, the model we were working on was quickly thrust into the spotlight. For example, in October 2022, we officially open-sourced the single-model multi-tasking large model PromptCLUE. At that time, there were not many downloads. However, after the launch of ChatGPT, especially in January this year, as the topic of ChatGPT continued to ferment around the world, The number of downloads of our open source models has grown rapidly, which is a relatively big boost for us and has also accelerated model development.
Second, talent. Previously, under the influence of the domestic atmosphere, many domestic and foreign talents joined the AI wave, but it is inevitable that some people are a little disappointed with this track. Now that this huge wave is coming, many talents have returned to the general direction of AI, so I think in the long run, this will be a major benefit to my country's technology industry.
Chen Shi of Fengrui Capital : Actually, I used to write programs, and I have a good understanding of AI technology and applications. I myself have also experienced the wave of AI in deep learning computer vision before, so this time I am doing generative AI, mainly pre-training large-scale language models, and GPT-4 has added multi-modal input and output. This technological change surprised us very much, because it was a sudden jump, not a continuous and gradual change.
The main reason behind this is that AI already possesses some general intelligence. Microsoft Research recently reported that we have seen the spark of general artificial intelligence in the current development of language models, and this spark is still burning and growing. This is a very surprising thing.
From an industry point of view, I think this wave of technology is very different from the previous wave of deep learning. The last round of AI technology reform did not meet expectations on the whole, mainly because of its poor versatility. In the end, there were only some real implementations in a few application scenarios such as security and face recognition, and the input and output were not good. . But in this round of AI technological change, the versatility of the technology is better, the applicable scenarios are wider, and the output value will be greater.
I think that if AI can improve the efficiency and experience of human society. And this kind of change can be seen, and it will also fundamentally change all walks of life, especially content production, education, scientific research, etc., and many other industries will be affected.
So we are very excited about this wave of AIGC. This is the well-being of mankind, but the premise is that AI does not betray mankind; this also brings great opportunities for entrepreneurs.
Unbounded AI Ma Qianli : From the perspective of entrepreneurs, this round of AIGC upsurge has made me even more anxious.
Because its development speed is too fast, we estimated in August last year that in February and March this year, the influence of AI on picture content will become stronger and stronger, and the changes will be reflected in comics, GIF, short videos and other fields. But unexpectedly, after Controlnet joined in the first quarter, a lot of crushing new achievements were born.
This can cause anxiety at times, because while you are doing a lot of work, other teams may catch up.
On the other hand, this kind of anxiety is actually industry-wide, and even people in traditional industries are also anxious.
Instead, it formed a kind of "huddling together for warmth". Some "unattainable" partners approached us, such as Chery Automobile, Ruotai Chaowan, and FMCG brands. They gave us their demands, and we worked together Do training, and even give a lot of privatized data.
You must know that under normal circumstances, companies will not easily share these data, but now they are willing to provide them, but they are also worried that if they do not participate in this wave, they will be eliminated.
Although there is anxiety, it is more of a feeling of progress in anxiety, which is a feeling of entrepreneurs.
Lin Laini of Huayuan : Let me talk about my feelings from the field of sapiens.
Digital Sapiens has upgraded from the past text customer service to voice customer service, and then to the development cycle of the multi-modal human-computer interaction system. The human-computer interaction mode is continuously evolving.
The current digital Homo sapiens technology has covered many areas of ability, such as answering questions, writing articles, text summarization, language translation and code generation, and even dealing with more niche topics; these multi-modal interaction capabilities In many business scenarios, Sapient has basically achieved "hearing clearly, understanding, and expressiveness", and is becoming more and more popular in finance, cultural tourism, media, public services, medical care, retail, entertainment, smart homes, etc. The commercialization of the industry has landed, providing customer service agents, financial advisors, broadcast hosts, tour guides, virtual idols, virtual singers and other auxiliary artificial services, which can be used as auxiliary tools for enterprises or individuals when completing some repetitive tasks.
Moreover, we all know that the world is still severely under-resourced in medical care and education. There is a huge demand in these two areas but there is not enough labor to meet the demand. This is precisely one of the main applications of Sapiens and even artificial intelligence in the future.
For example, it is very difficult for ordinary people in many developing countries to seek medical treatment due to their remoteness and lack of primary medical resources.
For example, set up some professional medical assistants to assist doctors in completing basic pre-examination and triage work, and provide patients with suggestions on whether they need further treatment and related precautions, so as to improve medical work efficiency and service levels; in education In the field of AI-enabled education, personalized learning programs can be provided according to the characteristics of different students, cultivate students' interest in learning, improve the quality of education, and so on.
Therefore, the emergence of Sapiens is both a challenge and an opportunity. We need to maintain keen insight and forward-looking thinking, and actively deal with its positive or negative impact.
How to reflect the innovation of China AIGC?
Qubit Jin Lei : The second keyword to discuss is model homogeneity and innovation.
From the current point of view, there seems to be a trend of homogeneity in both the gameplay of the AIGC application and the algorithm model behind it. Therefore, how should the originality of the Chinese AIGC be reflected?
Yuanyu Intelligence Zhu Lei : This question is very good, and we have been thinking about it.
It can be seen that most of the basic research, especially the algorithm model, was first born abroad. Looking at the current large-scale models in China, whether it is a large factory, research institution or a start-up company, it is very weak in the basic research link. Therefore, in terms of basic research or original exploration, the country still has a long way to go to catch up.
But this is not to say that the current domestic self-developed large-scale model has no value, on the contrary, it is very imperative.
First, this can promote the ecological construction of open source datasets and models.
For example, in terms of combing Chinese corpus, the quality of general corpus is not good enough; there is also a relatively poor atmosphere in terms of open source. As far as we know, there are many open source data sets in the medical industry abroad, but we only know 2 domestically. Open source data The quality of the set is not good either.
Therefore, more people should do this. This is also the direction and concept that our Yuanyu Intelligence core team was advocating when we launched the Chinese language model evaluation benchmark open source community in 2019.
Second, the domestic emphasis on large-scale AI infrastructure has reached a very high level. At this time, the attraction of talents, including many outsiders, are willing to join this wave. From a long-term perspective, this matter is very valuable for domestic AI infrastructure.
The innovation work of domestic large-scale models is being done by both major manufacturers and start-up companies, and each company has different entry points.
For example, there are specialized ones related to psychological counseling, such as metalinguistic intelligence, which is based on the general large model and makes an upper-level professional model. Focus on one or two industries and solve specific problems of customers.
For domestic startups, innovation is mainly reflected in whether they have a sufficient and deep understanding of the industry, and whether they can quickly run data and scenarios in the industry. This requires start-up companies to make certain innovations at the level of model underlying data and instruction fine-tuning.
Unbounded love for thousands of miles : There is no conflict between product innovation and embracing the Stable Diffusion ecology. SD is only helping to build a prototype from 0-1, but it still needs to innovate from 1-100 in terms of underlying data, middle-level algorithms, and upper-level applications. How to make this difference is something every entrepreneur needs to think about.
For example, we are now developing the national style model. First of all, the bottom layer of the model must have a large amount of national style data, such as Chinese faces, eyes, Hanfu and other materials. Taking Hanfu as an example, we can even specifically and finely distinguish the differences between different periods, different ethnic styles, and the left and right skirts of Hanfu. The meanings of different representatives are very different, and the details are ghosts and gods. For those who pay attention to the national style, the details are very important. Whoever trains this kind of scene model with rich details will be able to have the underlying differentiation.
The training data is not one-time, but real-time and self-iterating. Every day, 2 million users on our platform create close to 2.2 million pictures, and many pictures will be shared in our "square". Get real-time enhancement, so the differentiation is even more reflected. This is why we use the open source Stable Diffusion, but the platform works are very different from the content on the market. Real-time self-iterative data training has become an innovative point.
Of course, in addition to the underlying data, there are also many innovative points in the training method, such as the integration of GPT functions. I will not go into details about these functions. You are welcome to experience it directly with Unbounded AI.
Lin Laini of Huayuan : Sapiens, as one of the applications of generative AI technology, itself includes several levels of content. Just like when we evaluate an actor, we will start from the "sound table". In addition to external parts such as appearance and voice, the behaviors, text expressions, voice expressions, interaction methods and even personality of Sapiens are all within the scope of generation.
Mr. Ma just answered the question in terms of the appearance, type and style of Sapiens. I will make some supplements from the content generation of Sapiens. According to the calculation of Huayuan, there are three levels of Sapiens. The first level is the ability to speak and move, the second level is the real feeling, and the third level is the ability to think and think.
Among them, the first level is related to audio and video generation in AIGC technology, the second level is related to text generation and cross-modal generation, and the last level is that the robot can have its own thinking. After continuous independent learning, it involves strategy generation, etc. In order to achieve the third level of Sapiens, we add the social common sense in the "common sense knowledge map" to the dialogue interaction framework to realize the emotional system of Sapiens; by establishing a behavioral probabilistic knowledge map, the map covers 5000+ Different behaviors realize the personalized interaction of Sapiens.
Therefore, the final generated Sapiens may have similar voices and appearances, but the combination of their body movements, interactive content and differentiated personalities is enough to allow Sapiens to have their own emotions and personalities, and possess unique originality. Of course, at present, Huayuan still needs continuous and in-depth research in the field of cognitive intelligence to achieve this kind of Sapiens with multi-modal interaction capabilities and its own personality and emotions.
Chen Shi of Fengrui Capital : In terms of trends, not only domestically, but also overseas, there are a lot of homogeneous follow-ups. According to statistics, in 2022, an average of one large model will be produced every four days, and there will be more than 90 large models, mainly in the United States.
The core breakthrough of this round of generative AI is still at the algorithm level and model training paradigm level.
In terms of algorithms, everyone just mentioned Transformer and Diffusion Model, which are the current mainstream models, so many models now use them as the underlying architecture to iterate and change.
From the perspective of popularity, I don't think this is a short-term boom or bubble. I think it will go through a long period of progress and generate very large social and commercial value. Some people in the industry say that a large language model is a human-machine interface or a new form of operating system.
First of all, the language model has become the first interface of the human-computer interface. In the past, people could only make limited choices through menus and graphics, but in fact, language is the most natural and flexible choice for human-computer interaction.
In addition, from the perspective of the operating system, because language models such as GPT already have certain general intelligence, it is a bit like a brain or a distribution center. It can interact with humans, receive, disassemble and distribute tasks to various external plug-in applications to get results. Then feed back to humans, so it can be considered that it also has the characteristics of an operating system.
The software industry is also changing in the future. I think the direction is an application ecosystem centered on the language model. In addition to AI empowering and improving industry efficiency to generate commercial value, humans can also learn from AI to continuously improve their own learning capabilities.
For example, it is very painful for human beings to learn languages. For example, the effect of learning foreign languages for more than ten years may not be very good, but the efficiency of machine learning languages is very high. If we can partially open the internal structure of AI and find some of the laws, Perhaps humans can also use this to improve themselves and progress together with machines, and these things have great social value.
Therefore, I don't think the AIGC boom will be short-term. It has long-term social and commercial value.
Can a large model only rely on "strength to make miracles"?
Qubit Jin Lei : Regarding the parameter scale of the large model, there is another issue to be discussed, which is called "strengthening miracles". Dali will surely produce a miracle? I would like to ask Mr. Ma first, do you think that the method of vigorously producing miracles must be the only way for AIGC?
Unbounded love for thousands of miles : If AIGC can perform a miracle, no matter it is strong or what kind of force, it will be a good thing.
In some contexts, "strengthening miracles" seems like a bad thing, like a last resort choice, but it is not necessarily a bad thing.
I once read a blog called "Bitter Lessons", which was written by Richard Sutton, the father of reinforcement learning. He believed that human beings always intervene in the machine, teach it how to play Go, and let it learn human chess records, but in fact The most effective method is still self-learning, unsupervised learning.
To a certain extent, too much human intervention will limit its performance. So in essence, although we need a lot of computing power and follow up in the algorithm, we don't need to intervene so strongly, otherwise the effect may be completely different.
Chen Shi of Fengrui Capital : It is inevitable to make great efforts to produce miracles, because the breakthrough at this stage is achieved by unsupervised learning.
Unsupervised learning, especially large-scale unsupervised learning, has been the holy grail of machine learning until now, and no one has ever achieved this holy grail. Now that OpenAI has taken it down, I think this is something that has been verified at the moment, and it is also a trend, so it is estimated that everyone will continue to do this. If there is no major breakthrough in algorithm and training technology, I think the big model is still not enough. This is the only way to train.
For startups, if they are working on the application layer, they don’t need to train the model themselves, but do tuning on the basis of the large model. So I think that most of the opportunities for AIGC startups may still be in the application layer, or in the direction of some vertical models.
As for the large model itself, I think that in the future there may not be too many companies in China, or even too many companies in the world. But vertical models still have some opportunities.
How to view the Chinese version of OpenAI
Qubit Jin Lei : The next keyword is called "Chinese version of OpenAI". In this wave of upsurge, many bigwigs in the country have stepped down to make plans and spread AI hero posts widely, saying that they want to build a Chinese version of OpenAI.
So, judging from China's current environment, is it necessary to build a "Chinese version of OpenAI"? Is it feasible?
Huayuan Lin Laini : I have full confidence in the emergence of companies like OpenAI in our country that have developed epoch-making innovative technologies or applications based on underlying algorithm research, especially when I saw many young entrepreneurs of Generation Z on the scene. Spirit, but also feelings and a sense of responsibility.
But becoming a company like OpenAI does not mean creating a Chinese version of OpenAI . We don’t need to repeat the path of others. In the field of artificial intelligence, “innovation” is very important. Although OpenAI has created an underlying platform for multi-modal large models, which has opened the door to the application of generative AI technology for us, but from the perspective of real-world application scenarios, multi-modal large models are not "one-shot, eat all" sky".
For example, from GPT's 117 million to GPT-3's 175 billion, the number of large model parameters has increased by nearly 1,500 times in the past few years, and the number of Google's Switch Transformer parameters has reached 1.6 trillion. Computing power is something that many companies and even industries cannot provide.
Therefore, how to improve the robustness of the algorithm, improve the effect of the model, and achieve training in new fields with a small amount of data through small-sample learning and multi-modal learning are still issues that AI needs to consider.
At present, there are already many AI companies in my country that have exposed the sharp corners of Xiaohe and are thriving. Although there is still a gap with the world's top AI companies like OpenAI due to insufficient experience in computing power or algorithms. But I firmly believe that in the near future, we will see our own AI representing enterprises and their innovative technologies.
Yuanyu Smart Zhu Lei : I think this statement is a bit overgeneralized. OpenAI's so-called "ten years of grinding a sword" is because it was originally established as an organization for non-profit purposes. Recently, it has cooperated with Microsoft, and the whole has entered a closed-source state, and there are also some commercial models.
But in fact, many foreign companies are doing technical research aimed at commercialization, and there are many technical researches not aimed at commercialization in China, such as universities, research institutes, or some open source communities, etc. There are a large number of developers Developers and organizations are doing such things for the purpose of open source.
It’s just that under the current domestic environment, people generally only pay attention to some projects with a commercial aura, but people pay less attention to those non-commercial projects.
In fact, including investors like Mr. Chen Shi, or media like qubits, everyone may really want to do one thing, which is to promote the development of domestic open source data sets and open source models. This may become the future. a cornerstone of
I think there is no essential difference between domestic and foreign countries in this matter. Everyone has some companies that consider commercialization and organizations that aim at open source.
Expectations for new opportunities for AIGC in China
Qubit Jin Lei : The last key word is the topic of the roundtable forum "New Opportunities for AIGC in China". Let's talk about our expectations for it in a short sentence.
Lin Laini of Huayuan : Speaking of this expectation, we found that Internet companies may not be very friendly to people aged 35+, so we are also thinking about a question, will AI really create a wave of unemployment? Then I think not. I am also a middle-aged person over 35 years old, and I will talk about my feelings here.
I really like an ad slogan, which is "Never give up, run to the future". Although we are over 35 years old, we may also use some mechanisms to transform ourselves like Transformer, never give up and run to the future, use A positive and courageous attitude faces the opportunities and challenges of the future.
Unbounded AI Ma Qianli : I think it may be a more complicated one. Look at the recent news that thousands of scientists, including Musk, believe that AIGC and AI large models are a dangerous thing.
I think the domestic opportunity lies in that even if there is a gap between you and foreign countries, even if there is a gap of several times, it is enough for human beings, just like the Soviet Union and the United States during the war, they all said how many times they could destroy the earth, but in fact It is enough to destroy once. Maybe the AIGC's ability will be so strong that it doesn't matter what percentage difference there is between the two models, as long as it is enough for human beings. This is a long-term view.
So even though our current big models may not be as strong as foreigners, logically speaking, we can still catch up.
Chen Shi of Fengrui Capital : This opportunity is actually an opportunity for all mankind, and China also has a very big opportunity. At present, whether it is OpenAI, Google, or some other large foreign companies, many of our Chinese engineers or Chinese engineers are deeply involved in this research and development process. I think that China has the opportunity to make a large-scale basic model, especially a large-scale language model, which can be done. why?
At the algorithm level, this is actually a common scientific research achievement of everyone. From the pre-training process of GPT-3 to the reinforcement learning process based on human feedback used by ChatGPT, there are related papers detailing the algorithm and implementation logic. , most of the training data is also open. Although the details of OpenAI may be somewhat unclear, in fact, the most difficult thing left may be some specific engineering implementation work. I think our Chinese engineers are not afraid of specific project implementation.
In terms of computing power, I think this matter needs to be done slowly, try to think of some alternative methods, so as to have our own computing power, or obtain some computing power through other methods.
In terms of data, I think we can do it, because of the data set, it is now 450 billion Tokens (about 0.7 words each), and the data of this volume is relatively speaking. After our efforts, there should be some ways to collect and organize them.
I think that with the intelligence of the Chinese people, the basic model can be broken through. Of course, there may still be a gap to reproduce the level of GPT-4 today, but it is only a matter of time. Our application ecology will definitely flourish based on these foreign and domestic models.
In fact, including WeChat, DingTalk, or other mobile application ecosystems of the year, we have repeatedly proved that China is very powerful in software applications, so I think this is a new opportunity for China.
Zhu Lei from Yuanyu Intelligence : I very much agree with Mr. Chen Shi. Although there is still a technical gap compared with GPT-4, it is indeed a matter of time, because from a global perspective, except for Silicon Valley, it is the domestic boom.
The huge wave of AIGC in China is said to be the "iPhone" moment, but I think it is more like the early PC Internet moment. When everyone's awareness of AI is not so strong, a huge opportunity and opportunities suddenly appear. At this time, I think that for everyone on our stage and the audience in the audience, including the audience of the live broadcast, all walks of life , is a huge opportunity.
But at the same time as opportunities, this is also a challenge, so the last sentence is that we embrace AI.
Qubit Jin Lei : Thank you very much for your wonderful summary and sharing. Due to time constraints, this roundtable discussion ends here. We will reveal the answers to the new opportunities that the guests just looked forward to and look forward to, and how they will develop in the future.