Lu Qi has checked the large model project.
Author|Wu Jingjing
The clock of time has been crossed off halfway on the calendar for 2023. For many people in the technology industry, there is an illusion of being a world away.
OpenAI single-handedly drew a clear dividing line for the times - before the emergence of ChatGPT was AI 1.0, and after the emergence of ChatGPT was AI 2.0. Some people have begun to slowly get used to the new work-life model of the AI 2.0 era, they are accustomed to having new machine companions, whether it is asking for advice at work, upgrading work methods with new tools, or chatting with robots in their leisure time, telling the mind, the new relationship between machines and humans is opening the curtain.
New technologies unlock new possibilities for building applications, and entrepreneurs who dare to pursue their dreams will not give up this new technological opportunity, and many people can't wait to set off to explore the new possibilities of the AGI era. Some people say that ChatGPT has brought a new round of App Store events, and a variety of tools and products have been born, ChatPDF, various Copilot... Almost every day in Silicon Valley, dozens of new OpenAI-based products come online.
China's new technology startup market is also booming, and last Saturday, the Spring Startup Roadshow of Qiji Chuang Tan's event returned offline, and we saw the 60 most cutting-edge early-stage start-up projects and a group of enthusiastic entrepreneurs.
Zhu Xiaohu, managing partner of GSR Venture Capital, Zheng Qingsheng, partner of Sequoia Capital, and many partners of first-line funds also came to the scene. "Lu Qi's appeal in the field of AI is unquestionable," one investor said. The scene was hot, with more than 1,100 investors sitting in the audience, and many of them stood to watch the roadshow for more than two hours.
The average age of these founders is only 29 years old, some of them are sophomores, some are doctors and professors from the laboratory, some are CEOs of listed companies, they are attracted by the legendary leader of science and technology, Lu Qi, decided to put entrepreneurship on the scale of life, trying to turn one idea after another into a real business product. At the scene, we also saw that BabyTree founder Wang Huainan was also inspired by new technologies and brought new AI projects.
Lu Qi mentioned in his opening speech that this time the entrepreneurship camp received a total of 7954 applications, and finally admitted 60, with an acceptance rate of less than 1%, and 100% were technology-driven.
Looking at these 60 projects and backgrounds, "Jiazi Lightyear" found that these founders are basically from top universities at home and abroad, of which 78% have master's degrees or above, 67% have overseas backgrounds, and 32% have female founders.
The roadshow covers almost most of the subdivisions of science and technology, involving 22 sub-industries such as artificial intelligence and large models, enterprise services, robots, developer tools, biomedical, and advanced manufacturing, of which artificial intelligence projects have doubled.
Last year's metaverse and web 3 projects disappeared, and in their place, large models became the main play, with 39 large model projects out of 60 projects, accounting for 65%. This is followed by enterprise services and robotics, of which the number of robot projects has increased nearly 4 times. We also found that going overseas has become a hot spot, and the target market of many projects is not limited to China, among which developer tools, robots and other projects have taken overseas as the first stop for commercialization.
At a time when big models are hot, what new ideas and thoughts do entrepreneurs have? This time, "Jiazi Lightyear" shared some of the entrepreneurial projects related to large model technology that we saw on the scene for everyone to learn from.
1. Big model + software product = infinite possibilities?
The big model is like a magic wand, at this moment, every entrepreneur obsessed with technology wants to become a Harry Potter with magic, they create the ideas in their heads into lines of code "spells", and as soon as the spell comes out, the new product "magic" appears.
Yang Guang wanted to use the "magic wand" of the large model to customize the clothing, and as soon as the "spell" came out, a new dressing experience came.
Guang Yang is a designer who studied computer science and statistics at the University of California, Berkeley. He created his own personal clothing brand at the age of 18 and once designed clothing for Faye Wong and Sun Li. He said he hopes to allow more ordinary people to wear their favorite clothes without being limited by the established size.
Now, inspired by new technologies, Yang Guang and his partners Jin Di and Dai Chen have created a new brand "MohAIr Mohair", an AIGC personalized clothing customization platform, he and his team want to integrate technology and fashion to make high-quality, low-cost personalized clothing solutions possible.
MohAIr's product idea is: using keywords or pictures, users tell the system what type of clothing they want, the system can generate the corresponding design within five seconds, and let users intuitively see the effect of the upper body of the rendered clothes, and then, as long as you enter body data such as height and weight and place an order, the clothes can be delivered to the hands within a week.
Yang Guang, who wore clothes designed by MohAIr when he spoke on stage, revealed that the current internal test users of the product have exceeded 5,000, and the average daily design output exceeds 500 sets. It was learned at the scene that the current unit price of MohAIr clothes is about 3,000 yuan, and there will be hope to further reduce the cost in the future.
If Yang Guang's idea of "making new clothes" is eye-catching, then the idea of Blue River creating "external brains" sounds very fantastic.
Blue River said that in the future, people will need two AI partners: one is to help deal with various life things like Javis, the butler in the Marvel movie; The other is his own avatar, which can interact with other software on his behalf. They hope to use big models and innovative software and hardware to do these two AIs well, bringing a new lifestyle driven by big models.
An expert in materials science, Blue River graduated from Imperial College London and dropped out of the University of Manchester for many years in flexible electronic systems and wearables. His vision of the new intelligent hardware is: the robot's eyes, ears and other five senses are worn on the body, and connected to the external brain through nerves, so that the external brain has the real memory data of various ontology, and this wearable device is almost invisible and looks no different from normal clothes. The business plan of Exobrain Technology is to launch products in the third quarter of 2023 and sell more than 100,000 hardware products by 2024.
Like Lanhe, Wu Jingchao and his alumni from Zhejiang University also wanted to create an AI avatar, but they chose pets.
What their team has in common is a "love of pets" and a "passion for exploring AI", and they have created a pet AIGC product, DreamPets, "We want to create AI virtual avatars for 1 billion pets." Wu Jingchao said.
DreamPets' product concept is simple, users upload a few photos of their pets to get an AI avatar of their pet. Each user has their own fine-tune model, which can generate their own pet pictures and pet story content, and then the company will try to make different pets have their own personality traits and interact with the owner individually.
So far, in a 6-week limited test, DreamPets has gained nearly 100,000 users with an average weekly growth rate of 380% and has created 100,000 pet models.
In addition to creating various avatars with large models, various AI Copilots for different industries and life scenarios are also being born.
BabyTree founder Wang Huainan brought his new startup ChatBrook. ChatBrook's vision is to record the moments and lifetime of his/her life for each individual at the lowest cost and highest quality. Let everyone have a Sima Qian by their side.
Based on Wang Huainan's live talk, we can learn that this looks like a new community tool that can record and socialize. He mentioned that Ruxi's product planning has four levels: building a recording tool based on an intelligent big language model; Upgrade tools to become a recording community and do high-sticky socialization; Make recording and socializing the foundation of training digital twins; Build a digital twin application ecosystem MaaS (Memory as a Service).
Over the years, Wang Huainan has been a serial entrepreneur, and the mother and baby online platform BabyTree, founded in 2007, has been listed on the Hong Kong Stock Exchange. In 2021, he founded Mica Technology for silver-haired people, and in 2022, he launched a comfortable and fashionable footwear brand "Xiangwu", which is his fourth entrepreneurship. Miao Yuan, another founder of the company, was CEO and co-founder of MobLab, which created the world's largest behavioral gaming and artificial intelligence platform.
NetPrism hopes to create personal shopping AI Copilot, a project made up of University of Washington students and alumni, founder Wendell Li released the Restock Alerts inventory monitoring app during his studies at the University of Washington, which was popular in North America, ranking 7th in the App Store. He believes that scenario-based communication will become the next shopping entrance after e-commerce social networking and live broadcasting.
They designed a conversational shopping guide product that applies the large language model to the online shopping retail field, and through talking with users, BuyScout can put forward personalized purchase suggestions and product comparisons according to shopping needs and scenarios.
Users simply tell BuyScout, "I'm going to get a ski trip and the slopes are near Seattle," and it automatically filters out the best gear for skiing on Seattle's slopes. Similar to this project, Google's previous shopping experience optimized in search scenarios, but Google is to accurately recommend shopping and buy ads through content.
At present, BuyScout has 10,000 users, is the seventh shopping assistant in Apple's APP store, and their target market is 270 million online consumers in Europe and the United States, and domestic friends may not be able to experience it temporarily.
In addition to shopping, more tools for subdivided industry scenarios are also emerging, To0 Turing World to create an AI tool platform in the field of architecture, "With new large model technology, we hope to make the best quality delivery-level AI renderings on the market, and in the future, we will also launch a design interactive APP, so that everyone has a master AI designer." Founder Liu Xingchen said that he comes from Tsinghua University majoring in architecture.
In the To0 Turing world, users only need to simply enter design requirements or upload photos, click on the design template, and get high-quality renderings, shortening the original delivery cycle from one week to a few seconds, and reducing the cost of purchasing renderings from thousands of yuan to 100 yuan.
The above projects are only a small part of the native applications of large models, and new technologies are continuing to expand the boundaries of imagination, facing social, finance, designers, animation, and more diverse AI native applications are still emerging. We will continue to follow up the report in the future.
2. With the blessing of large models, how to make more intelligent robots arrive?
In addition to the innovative products of various software, there are also many people who want to put large models in the robot, so that the general robot can really enter thousands of households as soon as possible.
One of them, Dicong Chiu, who studied at Carnegie Mellon University's School of Robotics, has been exploring the cross-field research and implementation of AI and robotics in the past few years, and has also participated in the development of NASA's Mars rover. This time, he founded Jacobi.ai Jacobi Robot, which wants to combine multi-modal robots, AI large models, and cross-form motion control technology to empower all robot companies with RobotGPT and plug-and-play J-Box, so that intelligent robots can quickly enter all walks of life and thousands of households.
Its product thinking is: on the one hand, build a robot thinking brain J-Mind, which allows robots to speak directly, write instantly, and respond quickly according to different business needs; On the other hand, through the robot motion control cerebellar J-Box, this intelligent ability is plug-and-play, and the R&D deployment and operation and maintenance of general-purpose robots are realized with short cycle, low cost and high flexibility.
This is a team that integrates robots, large models, and multidisciplinary experience in chips, and Ye Gengen, the head of hardware chips and production lines, has served as a senior engineer at Intel, Qualcomm, etc., and has more than ten years of practical experience in the field of DSP, ARM, FPGA multi-core heterogeneity. Dr. Liang Junwei, a large model scientist, graduated from Carnegie Mellon University and later worked as a senior scientist at Google AI and Tencent Youtu Labs.
InsideRobo is also exploring the possibility of reducing the cost of robot intelligence by designing, training, and testing robots in the metaverse. It hopes to help enterprises reduce the cost of robot design, training, testing and management by developing real-time, realistic task flow simulations.
Their technical thinking is to use the latest metaverse simulation engine, through Nvidia Omniverse as a digital simulation and cloud computing platform, to create a high-precision, high-fidelity virtual world, and use large models and Internet of Things technology to synchronize robots in virtual and reality.
The founder, Dr. Yizhou Zhao, studied at the University of California, Los Angeles (UCLA), engaged in computer vision, robotics and computer simulation, and his supervisor was Professor Zhu Songchun. Another Gavin Steven, who works in cloud computing, IoT and industrial metaverse design, was a Microsoft Azure cloud architect and DevOps team leader.
The number theory technology is a general-purpose robot that wants to directly make entities, and they hope to create a generalized integrated intelligent industrial robot based on a large model.
Their product landing thinking is: directly understand and control industrial robot equipment such as robotic arms, AGVs, cameras, grippers and other industrial robot equipment through language, replace complex processes such as traditional annotation data + model training + programming teaching + simulation rendering, and realize the efficient replacement of robots in more flexible production processes.
This is a team from the NPC, and the founder Huang Zhen studied at the NPC with both bachelor's and master's degrees, and was the general manager of the Shenzhen branch of the AI company Qunzhi. Wen Jirong, Chief Scientist of Big Model, is the Executive Dean of Hillhouse Artificial Intelligence College of Renmin University and has rich experience in the research and development of multimodal large models. CTO Liu Zhanliang has worked in Microsoft, Baidu, Tencent, KLCII, etc., and has rich practical experience in large model projects.
Different from the previous few, Dynamic Microvision hopes to help intelligent manufacturing and intelligent robots through special bionic vision large model chips.
The intelligence of robots not only depends on the research and development of software and hardware, but also requires basic computing power support. The original intention of Liu Min and his team is to combine the vision model and advanced bionic eye vision sensor to provide users with a full set of vision solutions with ultra-low power consumption, real-time response and no privacy problems.
At present, MicroVision has created the first dedicated bionic vision processing platform, covering modules and chips, combined with vision models and advanced bionic eye vision sensors, to help intelligent manufacturing, consumer electronics, autonomous driving and humanoid robots and many other fields.
The founding team is from ETH Zurich, and the founder Liu Min is a PhD and postdoctoral fellow at ETH Zurich; Co-founder Dr. Dongchen Liang also graduated from ETH Zurich, and the company's principal consultant, Tobi Delbruck, is a professor at ETH Zurich, an IEEE Academician and a well-known expert in neuromorphic engineering and computational neuroscience.
In addition to the new attempts of large models + robots, there are also teams aiming at rehabilitation, cleaning, agriculture and other scenarios, trying to further open up market opportunities through innovative products and designs.
Founded by leaders from Google and Cobos, Miaoguo Technology chose to take multi-floor home cleaning robots as the starting point to create a comprehensive smart home service ecology.
The founder Long Yun is the former head of Google's TPU model optimization team and has deep accumulation in the implementation of AI solutions. Co-founder Wang Youdong is a former leader of the Ocean Lab robot project, and another co-founder Zhu Feng is the former head of Cobos North American sales system, and has led Cobos North American sales from tens of millions to hundreds of millions of dollars.
Aiming at the agricultural scene, Moss Robots chose to target the overseas agricultural market, using tree nurseries as the entry point, hoping to change the operation scenario of high-value crops by integrating intelligent automation and digital technology. Founder Yudan Hu studied at the California Institute of Technology, specializing in liberal arts studies on sustainable agricultural practices and labor issues. CTO Shrijit Singh has six years of experience in robotics.
The track chosen by serial entrepreneur Shi Yunlei is rehabilitation fitness robot, and he was one of the founders of minimally invasive medical robot.
The body robot he founded hopes to develop modular, cost-effective, human-machine integration rehabilitation robot products, use modular force control joints and fun game library to build a rehabilitation robot matrix, achieve full rehabilitation scene and full rehabilitation cycle coverage, and will expand to the fitness field.
The Chief Scientist of Rujing Robot is Professor Zhang Jianwei, an academician of the German Academy of Engineering, a pioneer of cross-modal learning in the interdisciplinary research fields of computer science, neuroscience and psychology, and has rich experience and resources in the fields of medical rehabilitation and elderly care services.
With endless new ideas and technologies, the robot industry is also ushering in new changes.
3. Infrastructure opportunities in the era of big models
Before big model technology brings huge changes to our working lives through a variety of products, how to make this costly and difficult technology more feasible is also a key question.
Therefore, at the scene, we also see that many people with technical backgrounds choose to target the infrastructure layer of large models, starting from computing power, computing frameworks, data and various software tools, and explore the possibility of making large model technology more economical and easier to use.
What HeX-AI is doing is building an easy-to-use and efficient big model computing framework, and they want to make big models affordable for all enterprises. They are from Peking University, and their founder, Fu Fangcheng, will graduate from Peking University's School of Computer Science in July this year, and previously engaged in research on deep learning algorithms and systems, big data processing and analysis, and distributed computing. CTO Dr. Miao Xupeng also graduated from the School of Computer Science of Peking University and later worked as a postdoctoral researcher at Carnegie Mellon University.
The opportunities they are targeting are exactly what many large model companies deeply feel: large models are too difficult to deploy and computationally inefficient. HeX-AI independently built a river map system, which covers full-link functions such as pre-training, fine tuning, inference, and deployment of large models, and solves the problems of low operational efficiency, high cost of use, and high difficulty in the construction of private large models. With the help of the river map system, it can be automatically parallelized to achieve the optimal distributed operation performance; and fine-grained resource management to make full use of computing power; At the same time, it can also adaptively perceive task load and cluster resources.
Low-bit, founded by Yang Haojin, who has worked at Alibaba and Huawei, also hopes to make large models ubiquitous through a new AI training framework. However, their technical thinking is to use the low-bit frame model calculation engine to bring a significant improvement in performance, so that the landing of large models on mobile devices can be realized.
Yang Haojin was the Beijing team leader of Alibaba AI Lab and Alibaba Video Cloud's Edge Computing Lab, and was also the Chief AI Scientific Advisor of Huawei Edge Cloud Innovation Lab and Munich Cloud Intelligence Lab.
The opportunity seen by Zhongke Jiahe is also at the level of the tool chain, but what they hope to do is to cross the ecological gap between domestic computing power chips and large model landing, and provide generalized, low-cost, high-performance computing power resources for the domestic AI large model industry.
At present, Zhongke Jiahe focuses on compilation technology, and is committed to the company's products as cross-brand and cross-model standardized AI software bases, including AI training and inference tools, Fine-Tuning tools and operator implementation, as well as integrated AI computing power solutions including hardware devices.
Xie Chen founded Guanglun Intelligent to target the pain points of large models at the data level, and they hope to accelerate the landing of large models through synthetic data solutions. Xie Chen believes that under the development method of AI 2.0, synthetic data is the most effective way to solve the shortage of large model data.
At present, synthetic data is indeed an industry hot spot, with the continuous development of large model technology, the available text data on the Internet may be exhausted in the future, and there will be no new training data available, and it is a big opportunity to solve the data problem in advance. Gartner said that in 2024~2030, 60% of large model data will be "synthetic data", and Silicon Valley giants such as Microsoft, Google, and NVIDIA are accelerating the layout of related businesses in the field of synthetic data.
Guanglun Intelligent hopes to deeply integrate generative AI and simulation technology to provide multi-modal, high-quality, large-scale, and low-cost synthetic data to bridge the data gap in the AI era. The founding team has rich practice in large companies at home and abroad, and the founder Xie Chen has served as the head of NVIDIA (USA) autonomous driving simulation, the head of Cruise (USA) simulation, and the head of NIO autonomous driving simulation. Co-founder Yang Haibo was Meituan's director of corporate affairs.
Zeabur, founded by Lin Yuanlin and his team from Zhejiang University, is a platform dedicated to helping developers streamline the service deployment process. They are a group of college students whose love and exploration of development have led them to discover this new opportunity for developers.
Lin Yuanlin believes that the era of large models is coming, and developers are facing new challenges: cleaning of data sets, debugging of prompt words, etc., in this case, developers need a tool to help them focus on development. With Zeabur, developers can easily deploy any service with a single click, regardless of programming language or development framework, and Zeabur can automate the deployment process with the most appropriate deployment strategy without worrying about infrastructure cumbersomeness. Over the past two months, Zeabur users have grown more than 20 times organically.
Also for developers, Limit-LAB hopes to become an AI app development platform that every white-collar worker will use in the future. Founder Zhang Runzhe has interned at Huawei, PingCAP, Starcoin and other companies, and has engineering experience in distributed computing, cloud architecture, and computer graphics.
Limit-LAB's current LimitHub product hopes to solve Prompt engineering, management, security and other problems, and their product is currently gaining a lot of attention in the open source community, and their plan is to cover all companies involved in Prompt development after 2023, when AI apps are rapidly booming.
From various software and hardware applications involving clothing, food, housing and transportation, to making technology more economical and usable, the era of big models is coming, we don't know what kind of products will be changed, and we are still exploring how to coexist with new technologies next.
Right now, these entrepreneurs have set off, they are not afraid of the fog in front of them, but choose to measure with their feet and take the next step.
What possibilities they will bring, let's wait and see.