In the past 2021, one of the most eye-catching things in the Chinese intelligent industry is the successful listing of SenseTime, which is known as the first of the "FOUR LITTLE DRAGONS OF AI", on the Hong Kong Stock Exchange, becoming the "first AI stock in China"!
For many "AI innovators" who are running, this is undoubtedly an exciting news - it makes people see that AI entrepreneurship is not a playful language that cannot be put on the stage, nor is it a "cold winter". The "FOUR LITTLE DRAGONS OF AI", which was founded in the early years, tasted in the process of climbing and climbing, not only the bitter fruit of opening up the industrialization of AI, but also the sweetness of the pioneer of AI landing.
After Yoshua Bengio, Yann LeCun and Geoffrey Hinton, the "Big Three of Deep Learning" won the Turing Award in 2018, it is widely believed that there will be no more major theoretical breakthroughs in the field of artificial intelligence in the next decade; on the contrary, the development of artificial intelligence will increasingly be reflected in the application of AI technology and its combination with the industry.
In other words, "how much money can AI make" has become a new focus. And with that comes more and more AI startups.
Compared with the older generation of post-80s entrepreneurs represented by the "FOUR LITTLE DRAGONs of AI", the new generation of post-90s young scientists have some differences in AI entrepreneurship.
Let's say the investment climate. The older generation frequently catches up with the AI investment boom, and with the emergence of problems such as the sudden emergence of bottlenecks in deep learning and the difficulty of transforming AI technology achievements, the new generation of entrepreneurs are increasingly encountering investors' doubts and defenses in finding money, and are also facing more and more competitors.
In terms of technological development, the original AI landed on perceptual intelligence, and later AI merged more fields of knowledge (graphics, statistics, reinforcement learning, evolutionary computing, etc.), and the entrepreneurial direction of a new generation of scientists began to be more diversified. In addition to computer vision, there are speech recognition, natural language understanding, graphics, chips...
To this end, we have curated the "AI Innovators" series, inviting the younger generation of AI entrepreneurs to share their entrepreneurial stories with us. The first entrepreneur, a young post-90s scholar who is well-known in academia and industry, is Yang Zhilin, co-founder of Recurrent AI and Transformer-XL and XLNet.
1. "Radical" AI landing method
"I want a more radical and radical path to break down the barriers between academia and industry." When asked about the reason for starting a business, Yang Zhilin replied.
In the eyes of investors, Yang Zhilin, like other founders of Circular Intelligence, has a glamorous resume that can definitely be handed:
He studied in the Department of Computer Science of Tsinghua University in Crouching Tiger, Hidden Dragon, studied under IEEE Fellow Tang Jie, and graduated in 2015 with the first place in his grade;
Subsequently, he studied natural language processing (NLP) at the World's #1 Institute of Linguistic Technology (LTI) at Carnegie Mellon University, where he studied under renowned scholars Ruslan Salakhutdinov and William Cohen;
During his Ph.D., he collaborated with Turing Award winner Yoshua Bengio to publish the "Hotpot Q&A" dataset HotpotQA, and published XLNet and Transformer-XL as a work that had an important impact on the field of NLP, becoming one of the highest cited papers in NeurIPS 2019 and ACL 2019, and the number of Google Academic citations directly exceeded 10,000...
Photo note: Yang Zhilin's Google Scholar paper is cited on the homepage
Generally speaking, doctoral students majoring in computer science at Carnegie Mellon University often have to go through six years of study to graduate, while Yang Zhilin only took four years (2015-2019) to leave CMU and once became a prominent figure in academic circles.
As a leading young AI scholar, Yang Zhilin is determined to promote the large-scale application of artificial intelligence technology in real life.
Generally speaking, the way for young doctoral students to participate in technology landing is to enter a large factory with abundant funds and cattle people, and find an official position in it, such as his two doctoral supervisors Ruslan Salakhutdinov and William Cohen, who are the head of Apple's AI research and Google's chief scientist in addition to academics.
However, Yang Zhilin believes that the model of "scientists joining big factories" has limitations in the organizational structure, which does not allow him to participate more deeply in AI landing, nor can it fundamentally solve the bottleneck of AI landing in the industry:
"I think the common problem facing the AI industry is gap between academia and industry. We see teachers have some titles in industry, but in essence they're still doing research. Although basic research is important, it is impossible to break down this barrier, and there are still many steps between the research content and the actual landing."
Specific performances are: First, college teachers tend to be more inclined to academic research, have less contact with the industrial community, and lack the thinking and driving force of the industrial community; second, at the same time, although many Internet manufacturers will also recruit outstanding scientists to solve technical problems, their primary starting point is to empower business, rather than promote AI landing.
In terms of the organizational structure of the operation of the big factory, these outstanding scientists do not have enough resources or rights to promote the landing of products. The company's business direction will be adjusted, even if scientists are willing to promote the landing of a product, the cost and cost will increase significantly, and the conversion rate and efficiency of technology will also be affected by the organizational structure of the enterprise.
In Yang Zhilin's view, this is a very big restriction, which also causes that in large factories, the landing cycle of many AI technologies is very long and not agile enough. Therefore, when he graduated with a doctorate in 2019, he rejected the high-paying offers of big manufacturers such as Google, Facebook and Huawei, and chose to return to China to start a business.
Photo note: Yang Zhilin with two PhD supervisors Ruslan Salakhutdinov (far right) and William Cohen (far left) "The advantage of entrepreneurship is that we can decide the organizational structure of the company ourselves. Life is short, energy is limited, optimize the company's organizational form can effectively reduce intermediate losses, narrow the distance between technological transformation and social value." Yang Zhilin said.
Yang Zhilin recalled the AI technology review that his undergraduate and doctoral supervisors attach great importance to the actual value of technology, which has brought him great inspiration. The difference is that he will pursue the results of the landing more aggressively and go deep into the business to conduct research. His plan is to conduct academic research and technology implementation at the same time, and achieve results at the same time.
Circular Intelligence was established in 2016. That is to say, Yang Zhilin has been engaged in academic research and entrepreneurship since the second year of his doctorate.
It is precisely because of the embarrassing situation of the "big factory scientist" that he has always stressed that "we need new thinking". In Circular Intelligence, he is not only the head of AI technology, but also the product manager, which is undoubtedly a "paradigm-level innovation" and the best way to effectively break the barrier between technology and value in his mind:
"On the one hand, we will do basic research, such as pre-training, multi-mode, etc.; on the other hand, we will also go to the ground. These two things can enhance and promote each other."
2. Research and landing, two-wheel drive
The first three founders of Circular Intelligence, Chen Qicong, Yang Zhilin and Zhang Yutao, met in the knowledge engineering laboratory of Tsinghua University and all have a passion for "creating social value with AI". All three are from technical backgrounds, and later because of business development, Circular Intelligence introduced another "fourth leader" who is good at technical product operations in 2018.
Since its inception, Yang Zhilin has been the core technical backbone of the team. In 2016 and 2017, he and Chen Qicong and Zhang Yutao began to explore the application direction of technology.
For the new generation of AI entrepreneurs, 2017 is an important time node.
That year, the Google team proposed the Transformer model in the article "Attention is All You Need", which did not use the time series structure of the volume machine network and the previous RNN, and used the coding mechanism, and the encoding end contained both semantic information (Multi-Head Attention) and positional information (Positional Encoding), which could be computed in parallel, which greatly improved the training speed of the language model.
For entrepreneurs who focus on technology landing, this is undoubtedly a good news, which can reduce the time of pre-training, save research and development costs, and accelerate the matching speed of technology and scenarios. The emergence of Transformer has broken the monopoly of computer vision in the AI startup circle with deep learning, and has made a large number of startups based on NLP technology begin to emerge, and circular intelligence is one of them.
The main business of Circular Intelligence is to use artificial intelligence technologies such as NLP, voice, multi-mode, and big model to create "sales technology" solutions to help the sales team of enterprises improve sales performance.
Yang Zhilin said: "We believe that the process of AI generating value can be divided into several stages, and one of the stages is to help everyone become better, improve people's capabilities, and thus improve the operational efficiency of the whole society. This Vision was an idea that our company had at the beginning of its founding."
According to him, circular intelligence chose to use AI to "improve people's communication skills", which was also determined after a long period of exploration, communication with customers, and continuous iteration. In the end, the driving force behind their choice is the customer's demands and the judgment of the overall market. For example, according to CB Insights, the amount of investment in selling technology startups exceeded $5 billion in 2016, and has gradually increased since then. This also shows the market's confidence in this track.
Gartner's SalesTech technology maturity curve 2021 shows that sales empowerment has passed the "Innovation Trigger" and entered the "Peak of Inflated Expectations"
As mentioned earlier, Yang Zhilin believes that the organizational structure of the AI system will affect the capabilities of the product, and the flexible setting of the organizational structure can help them promote the product landing in a better mode. In the process of entrepreneurship, Yang Zhilin realized the two-wheel drive mode of academic research and industrial landing. For example, he was received by ACL 2019 transformer-XL article, which was used in as an ASR product for circular intelligence long before it was published.
For the rapid transformation of this basic technology, Yang Zhilin is proud:
"In the process of pre-training, we deploy the technology on the product system so that it can drive the intermediate research and development process with the ultimate goal of the actual data set as the ultimate goal. When the excavation system is landed, the system is also learned and optimized based on the final business results. At the same time, the intermediate process can iterate out a lot of AI problems and basic technologies, so that the products can be further improved later."
In the landing of AI models, a common problem is the authenticity and completeness of the data set.
In general, researchers often improve models based on specific, artificially created data, but these data may not be able to fully and correctly describe the situations encountered by the model in real scenarios. Therefore, although several pre-trained technologies have been tested on many datasets in academia and perform well, in practical applications, there are still many technical improvements that are needed to be deployed, because the model encounters more and more complex problems.
At present, there is still no outstanding progress in the academic community in resolving this issue. However, in the process of entrepreneurship, because Yang Zhilin and the team's pre-training technology research was tested in the actual data set from the beginning and directly matched with the AI product framework, similar landing problems can be uprooted from the root.
3. Talk about "NLP+ Sales"
In addition to the 2017 transformer, in recent years, many large-scale pre-trained language models based on transformers have emerged in the field of artificial intelligence, such as Bert and GPT-3. In addition, there are many emerging technologies that have had a positive impact on AI entrepreneurs, such as research breakthroughs with few samples and zero samples.
From the perspective of NLP technology landing, this will be a revolutionary moment. Because by organically combining these research results, the effectiveness and efficiency of AI models can be greatly improved. In some scenarios, researchers even need to use very few samples, or even zero samples, to achieve the same good results as ever.
For circular intelligence, this means that in the process of using AI to increase sales conversions, they can do a lot of things that they couldn't do before, such as conversational insight and analysis engines. With the gradual maturity of NLP landing technology, coupled with the fact that enterprise services have become a new investment hotspot on a global scale, Yang Zhilin and his team hope that through the path of "NLP+ sales", circular intelligence may also become a "Chinese Gong.io".
Specifically, the process of using artificial intelligence to improve sales efficiency can be divided into three steps: one is to collect the conversation data between sales and customers; second, to mine and model valuable conversation content, turning unstructured data into structured data; third, analyze conversation data, find out the problems that salespeople have in the process of communicating with customers, more accurately analyze customer wishes, and give key elements to solve problems.
Ultimately, it's about efficiently analyzing large-scale text data.
It sounds simple, but in fact, "AI + sales" is a track with the characteristics of both market needs and high technical barriers, because this requires AI systems to have the ability of comprehensive analysis, in addition to algorithms, conversational insight capabilities, data analysis capabilities, industry marketing knowledge and so on are also indispensable. This is very much in line with the entrepreneurial style of the circular intelligent founding team: not only to create value, but also to have a certain technical threshold, improve the difficulty of competition, and reduce opponents.
From the perspective of just demand, performance growth is the foundation of every company's development. Sales is part of marketing, and the impact of the conversion rate of its conversation traffic on business goals is crucial. According to Yang Zhilin's observation, traffic conversion rate is a prominent pain point in many industries, especially in the financial industry.
They had contact with a domestic head insurance company X, which had a branch Y. Y's sales dilemma is that although Z's sales team has sold many policies, much higher than Z, another branch of X in a neighboring city, the total premiums charged are lower than Z's. After analysis, the reason is simple: because the average premium of Y is much lower than that of Z.
At this time, they need to analyze the data they have in their hands to find a solution to save the sluggish performance. Looking back at the existing data, the only magic weapon for enterprises is to save a large amount of communication voice or text data.
To a computer, undecoded speech data is like a black box and unstructured. At this time, the combination of NLP and voice technology can efficiently parse this unstructured data and manage the communication process of an enterprise sales team. In other words, at this time, AI products are still a "management gripper" role, analyzing the communication between salespeople and customers, gaining insight into customer needs, and improving sales management capabilities and sales team work efficiency.
"We offer products that present the opening rate of each salesperson every day in a very clear and precise way. It can locate every phone call and every communication of every team member, so that you can do a lot of report analysis to know what the problem is with each team." Yang Zhilin introduced.
Figure Note: Schematic diagram of the principle of circular intelligent AI products
Based on the results of the actual battle, the AI system built by Circular Intelligence can handle more than 100 million conversations per day, helping Y increase the average premium of the policy by about 20%. At present, they have cooperated with dozens of enterprises with more than 1,000 sales personnel, mainly covering four major industries such as banking, insurance, real estate, and automobiles.
Yang Zhilin explained: "These industries have a common feature, that is, the requirements for sales skills are high, and the sales process is very complicated. At the same time, relatively speaking, these industries have the need for refined operations, and their degree of refinement has reached a certain threshold, which is enough to support them to apply AI systems to optimize efficiency." Leifeng Network
In this process, their AI system background has also accumulated thousands of semantic models from different industries, forming a powerful knowledge base of circular intelligent AI brains, which is conducive to the further landing of NLP models. Last year, they cooperated with HUAWEI CLOUD to develop a large-scale Chinese model "Pangu", which achieved an effect superior to the Bert and GPT series in some actual scenarios.
At present, Circular Intelligence has carried out B round financing, achieving more than 200% revenue growth for three consecutive years. However, Yang Zhilin said that entrepreneurship must go deep into the business: "We are still in the stage of product polishing and maturation, and the main task is to expand and improve the coverage of sales communication scenarios."
4. Entrepreneurial impressions
Yang Zhilin believes that the two factors that determine whether a technology startup can stand on are that it has multi-dimensional comprehensive capabilities, and the other is that it can cultivate the industry deeply and can organically combine general products with subdivided industry solutions:
"When we have a generic technology brand, we can expand it to new industries, new companies and new segments at a lower marginal cost. Therefore, we need a team that becomes an 'industry expert', who can provide a professional industry solution, and then use this professional solution to package products and land them." Leifeng Network
Loop Intelligence has a team of star founders, which is not difficult in attracting comprehensive talent. Yang Zhilin also stressed that creating ai systems to boost sales cannot rely solely on the power of a group of "NLPer" or "CSer", but also requires hardware talents, marketing talents, industry analysts and so on.
In the process of technology empowering the digital economy, the core value of NLP is often reflected in the last mile. For any industry, as long as there is a scene of communicating data and text data, NLP technology can play a value. The biggest bottleneck of the traditional NLP scenario is scale, but with the breakthrough of research such as Transformer and less sample/zero sample learning, which releases extremely high marginal value, Yang Zhilin believes that in the next few years, the large-scale empowerment of NLP will become possible.
In the five years of entrepreneurship, Yang Zhilin summed up his growth outside of technology: first, he had the opportunity to learn business logic and deepen his understanding of the industry and scenarios; second, he learned how to build and operate a company; third, he could have a more thorough way to shorten the gap between technology and value.
Ten years ago, perhaps many people would have thought that entering a big factory was the best way to study practical AI. However, with more and more technology bulls leaving the Internet factory in the past one or two years, or returning to academia, or starting their own businesses, people have begun to realize that to promote the large-scale landing of artificial intelligence technology, a new mode of operation is needed. From this point of view, Yang Zhilin's choice is quite prescient.
It is understood that at present, in addition to starting a business in circular intelligence, Yang Zhilin has also led a number of AI research projects in Tsinghua University, Zhiyuan Research Institute and other institutions, and continues to practice his thinking on how to break the barriers between research and application. Driven by the two-wheel drive of entrepreneurship and academia, a new generation of doers such as Yang Zhilin have brought not only "radical" applications to the future of Chinese intelligence, but also cultivated young talents who have begun to think about technological transformation from the research end.
Pass on the torch, the future can be expected.
Reference Links:
1. https://www.gartner.com/en/documents/4004056/hype-cycle-for-crm-sales-technology-2021