“Chinese-Style” OpenAI’s “Crazy” 200 Days: Building Models With PPT While Struggling to Find Use Cases

Image source：Visual China

BEIJING, June 23 (TiPost) — On November 30, 2022, ChatGPT, an artificial intelligence (AI) chatbot product developed by OpenAI, was released.

OpenAI probably did not expect that ChatGPT, originally a product that showcased the GPT’s capabilities to consumers, would attract widespread attention from investors, entrepreneurs, unicorns, big companies, academia, economists, and even the Minister of Science and Technology of China in the past 200 days. Meanwhile, more than 30 tech giants, startups, and institutions such as Google, Microsoft, and Alibaba have entered the fray, sparking discussions and a global AI model “arms race”.

The “China Large AI Model Map Research Report” shows that as of May 28 this year, 79 large models with more than one billion parameters have been released in China. The US and China account for over 80% of the total number of large models released worldwide.

A basic consensus in the industry is that the emergence of ChatGPT marks the beginning of general AI and the turning point for strong AI. It is a significant breakthrough in the field of AI technology innovation and application results, and also a “power plant” for digitalization in the new era.

With the help of ChatGPT, all digital systems and industries are worth redoing in a SaaS (Software as a Service) service manner, which can be accessed by various industries. In the future, more people hope that ChatGPT can make enterprise digital business processes faster, more efficient, and smarter.

However, compared with the commercial cases of ChatGPT used by companies such as Morgan Stanley and Stripe, which were announced by OpenAI and Microsoft, a strange phenomenon has emerged in the domestic “large model war”:

Although the technology and product capabilities seem strong, various bugs appear when the product is in the hands of customers. Companies that disclose large model dialogues are all talking about their own strong technical capabilities and scenario-based solutions, and some even disclose cooperation information, but they rarely talk about the process of commercial landing cases in public settings.

An industry insider revealed to TiPost App that a listed company complained during a phone conference about using a multi-billion AI model product developed by a certain internet giant. Despite claiming to be the first product of its kind among global giants and capable of creating PPTs in 3 minutes, the listed company experienced a “crash upon connection” when integrating the large model into its system.

At a recent AI industry application forum in Shanghai, an AI large model entrepreneur even stated that the dozen or so large language models released in the past few months are all very similar. The current situation is that only OpenAI can achieve commercialization of general AI and has the majority of users in the market. Furthermore, domestic AI large language models have yet to reach a commercializable level.

PPT-style large models can seve a variety of industries, but have numerous vulnerabilities when applied commercially

“In my opinion, whether ChatGPT can truly develop in this AI 2.0 wave depends on whether it has a business model and whether customers are willing to pay. No matter how we train large models like GPT, if there are no applications, no scenarios, no paying customers, and no business models, they cannot succeed.” On June 2, at a roundtable in Shanghai Lingang, Yao Wei, Executive President of Capitalonline, a cloud computing listed company, stated that commercialization of AI large models is extremely important for industry development.

From the customer’s perspective, enterprises also urgently need generative AI to bring business transformation.

According to a survey of 1,003 small businesses across the United States conducted by entrepreneur service platform GoDaddy, ChatGPT is the most widely used generative AI product in small businesses, with a usage rate of 70%; 38% of respondents have tried generative AI in the past few months; marketing, content creation, and business advice are the top three use cases for generative AI in enterprises; 75% of respondents are very satisfied with the performance of generative AI in their business.

As overseas large models are rapidly changing, OpenAI has launched a pilot subscription service for ChatGPT Plus at $20 per month, and ChatGPT/GPT-4 has opened APIs to developers with prices falling, driving the continuous landing of large models in application scenarios.

Meanwhile, many domestic large models have also been released. Based on the $200 billion generative AI market prospect and the $50 trillion digital economy industry scale, AI large models are expected to have the largest commercial development range in China.

Kai-Fu Lee, Chairman and CEO of Innovation Works, once said that the AI 2.0 era has entered a period of explosive productivity improvement applications, which presents a huge platform opportunity and is China’s first opportunity to participate in platform competition in the AI field.

Specifically, combining information on the development of large-model related enterprises, stock research reports, and Microsoft’s recent disclosure of application scenarios, TiPost App has summarized the main commercial applications of ChatGPT-type products in the following seven industries:

• Enterprise Operations: Daily office document writing and organization; marketing conversation robots, market analysis, sales strategy consulting; drafting legal documents, case analysis, legal article sorting; human resources resume screening, pre-recruitment, employee training.
• Education: Assisting in evaluating student learning situations and providing advice for career planning; customizing learning content based on student situations and interests; initial draft and review of papers; helping low-income countries/families obtain equal education resources through GPT.
• Gaming/Media: Customized games, dynamically generated NPC interactions, customized plots, open-ended endings; content generation for overseas copywriting, language translation and assisted advertising delivery and operation; live streaming of digital virtual humans; game platform code restructuring; AI-generated replicas.
• Retail/E-commerce: Monitoring and analyzing public opinion, complaints, and emergencies; writing and delivering brand marketing content; automated inventory management; automatic or completed SKU category selection, quantity, and price distribution; customer shopping trend analysis and insight.
• Finance/Insurance: Personal financial advisor; summary and initial approval of loan information; identification and detection of fraud risk activities; customer service center analysis and content insight; insurance claims processing and analysis; investor reports/research report summaries.
• Manufacturing/Automotive: Production plan and supply chain plan status queries; predictive maintenance assistance for production lines; product quality analysis and traceability; full-scene simulation training for autonomous driving and virtual car assistants; online car brand and configuration comparison analysis.
• Life Sciences: Target discovery and drug efficacy in the research and development phase; medical literature content retrieval, key summary extraction, and relevant regulatory sorting; medical representative training and knowledge base establishment; triage guidance assistant, diagnosis and treatment assistant, postoperative care and rehabilitation assistant.

Furthermore, ChatGPT’s large model and generative AI technology will also be extended to various complex scenarios in the fields of images, videos, digital humans, etc., using massive data resources and algorithms to achieve commercial applications and iterative updates.

OpenAI once conducted a study estimating that at least 50% of job duties in 19% of American jobs would be affected by AI. At least 10% of job duties in 80% of jobs would also be affected, with positions such as mathematicians, accountants and auditors, news analysts, legal secretaries and administrative assistants, and tax preparers being the most susceptible to the impact of large-scale GPT models.

However, the above content is mostly just theoretical, with PPT presentations imagining the intelligent upgrading of AI frontier technology. Although there is much competition in parameter scale during implementation, there are still few large-scale industrial deployments, and subsequent model corrections and iterative evolution progress slowly.

Whether it’s data presenting “nonsense” or inaccurate translations between Chinese and English, insufficient computing power, or high prices, the idea that AI will fully assist with shopping, finance, and manufacturing is one-sided. Submitting to customer applications is not an easy task and may result in conflicts and problems.

For example, at a press conference held by iFlytek, when asked a professional medical question, “What drug should be used to treat closed-angle glaucoma?” ChatGPT responded with atropine. TiPost App also attempted to answer this question, but the responses were mostly incorrect or not drugs that can be purchased under Chinese drug regulations. The correct answer is pilocarpine.

Additionally, because large-scale models often use English data instead of Chinese internet data, language barriers could arise. For example, when “fish-flavored shredded pork” is inputted, an awkward image of a live fish cut into shreds may appear, resulting in some problems during commercialization.

A senior executive at a financial company previously told TiPost App that due to ChatGPT’s poor math calculation abilities and inability to update some information in real-time, domestic large-scale model products are not very effective in the financial sector. There may be errors in exchange rate and loan information, and information asymmetry may also occur.

At the April release conference held by AI company Fourth Paradigm, a representative from a bank mentioned that information asymmetry in the financial industry could lead to higher interest rates for credit products or deposit rates, but even if all information is provided, the choices made may not be optimal.

“We are in finance, serving the public, and any information I convey must be accurate.” The aforementioned representative believes that the challenges facing large models in enterprise landing mainly include content credibility risk, data security risk, and high landing costs.

In manufacturing, content issues may have a more serious impact. Because human defects and errors are strictly limited, some high-precision processes require accuracy to the slightest degree. Once an AI system makes a mistake, it may cause an accident.

Zhou Bowen, Huixian Lecturer at Tsinghua University and founder of AI company XuanYuan Tech, told TiPost App that large models like ChatGPT are progressing rapidly, but the problem is that they may be seriously “talking nonsense”. Especially in professional fields, outsiders see it as an insider, and insiders see it as an outsider. At the same time, the original author of the content regards it as plagiarism, but ordinary users regard it as creativity. In fact, it does not yet have the originality of ideas.

In addition, language issues also need attention.

According to Wired, at least 15 arXiv research papers this year have explored the multilingualism of large models. However, researchers found that AI systems, including ChatGPT, are better at translating other languages into English, and it is difficult to rewrite English into other languages, especially Korean and non-Latin scripts. Moreover, ChatGPT performs poorly in answering factual questions or summarizing non-English complex texts, and is more likely to fabricate information.

At a US congressional hearing held in May, OpenAI CEO Sam Altman said that the ChatGPT development team is taking measures to narrow the language gap. He hopes to work with governments and other organizations to obtain datasets to enhance ChatGPT’s language skills and answer correct content.

Little mention of business cases, and some large models encounter obstacles in application

Industry insiders told TiPost App that GPT has already achieved true intelligence, and the next successful point is in the productization, commercialization, engineering, and application scenarios of large models.

According to a report by Zhuo Shi Consulting, the global AI market is expected to reach $199.7 billion in 2022, with a compound annual growth rate of 29.4%, and is expected to reach $562.4 billion in 2027, with a compound annual growth rate of 23.0% from 2022 to 2027.

“Today, the AI technology capability is vastly different from 5 months ago. We have put a stronger product on our system platform, and as for sales and service, from a business perspective, it has just begun. It takes time from exposure to new technology to final procurement,” said Huang Wei, founder and CEO of iFlytek, to TiPost App. The large model has just been released, and there are no large-scale commercial cases yet.

The scarcity of commercial cases is an important feature of the current domestic large model trend in AI. Even the AI industry giant, SenseTime, recently disclosed a figure of only 10+ large model customers, most of which are not vertical leading companies, according to TiPost App.

On May 30th, the generative AI (AIGC) company, UM Question, submitted its prospectus to the Hong Kong Stock Exchange.

The report shows that in 2022, UM Question’s total revenue will be 500 million yuan, and the top five customers are basically AIoT (Internet of Things) companies, most of which purchase UM Question’s smart IoT solutions such as smart watches. The customer cooperation has exceeded three years, rather than its AIGC business services.

As early as late April, UM Question announced the launch of the AI large model “Sequence Monkey” and began internal testing and exploration. Li Zhifei, founder and CEO of UM Question, told TiPost App that UM Question does not need external financing to support R&D investment. Apart from large models, the company’s other businesses do not burn much money. “(Large models) may be the last thing I do all in,” he said. However, TiPost App did not see specific commercial revenue from large models in the prospectus.

This means that the AIGC industry, including the ChatGPT large model, is currently unlikely to generate large-scale revenue.

Of course, in the past few years, when AI technology has been applied to landing scenarios, commercial customers have also been difficult to disclose their identities. This is not because the customers are in a highly confidential industry, but because AI is too widespread, and everyone wants to ride the “digital economy” wave.

For example, in October last year, TiPost App visited a leading sewing machine equipment manufacturer in Taizhou, Zhejiang, which is the world’s top sales company. The AI technology leader, SenseTime, served as the company’s logistics machine equipment and AI technology supplier and cooperated to build an “intelligent intensive warehouse.”

After TiPost App visited the actual commercial use of AI technology, they were told that the manufacturing leader did not want to disclose their cooperation with Megvii Technology, not because of the disclosure requirements of the listed company, but because the chairman of the company wanted to claim that it was their self-developed AI technology application, not Megvii’s. According to their disclosure on the Shanghai Stock Exchange, they raised over RMB 1 billion to invest in building an intelligent factory.

Therefore, when TiPost App talked to Megvii Technology CEO Yin Qi about this in March this year, he admitted that the term AI has been overused by everyone, and the popularization of AI has made it difficult for AI companies to promote themselves. Some things were talked about by others three years ago, so “we need to accept the current situation.”

“I don’t mind if our customers don’t say it was done by Megvii. For example, Huawei has done so much for operators, but in the early days, everyone just knew that Huawei was a great company, not what they were doing. Later, Huawei also made To C products, so they began to do some brand promotion. If the customer is willing to say it, it means that they also think this is important from this perspective. China needs some topicality for the capital market, and even some O2O companies say AI, which shows that whether it is related or not, everyone is doing AI.” Yin Qi told TiPost App.

Founders of several traditional manufacturing companies told TiPost App that they want to cooperate with AI technology R&D companies that do large models, mainly because startups are limited by high costs such as computing power, data, and electricity, and cannot refine large models on their own. By collaborating with AI company teams on data, they can learn the technical secrets and then build their own teams.

“We will never work with an AI company for a long time because the price is too high,” a supplier told TiPost App.

According to the prospectus released by Fourth Paradigm in April, although the company’s YoY growth in 2022 was 52.7%, its customer sources are quite diversified, and there are almost no repetitions among the top five customers in the past three years. The customer base is not very fixed.

Fourth Paradigm founder and CEO Dai Wenyuan explained to TiPost App that changes in the top ten clients do not mean that clients are changing every year, and customer retention rates are high, sometimes even reaching 90% per year.

How can companies address application of large models?

To solve the problem of large model implementation, there are mainly three aspects: improving content credibility; solving the problems of high computing costs, repeated training, and limited resources; the price of large models needs to be continuously reduced or vertical domain models can be used for implementation.

The first issue to address is improving content credibility.

Zhou Bowen, in regards to the TiPost App, stated that we should develop a universal large-scale model that can solve practical issues for different users. It should continuously apply feedback through commercial delivery and even require evaluation to solve credibility issues regarding content.

Zhang Bo, an academician of the Chinese Academy of Sciences and Honorary Dean of the Artificial Intelligence Research Institute at Tsinghua University, believes that ChatGPT lacks the ability to self-learn, which is the most fatal flaw of ChatGPT. Therefore, more data needs to be optimized to further address practical application issues.

“Do not assume that ChatGPT can solve all AI problems. Without the ability to re-learn, it cannot cope with changes. This is the same for both domestic and foreign ChatGPT. When I asked American ChatGPT, they gave the same answer. Some Chinese ChatGPT models perform well, while others are incorrect. This raises a significant question for us. We need to apply it to decision-making problems, which requires further resolution,” said Zhang Bo.

Xu Qingcai, head of the logistics business unit at SenseTime, mentioned in a recent exchange that currently, large models need to move towards verticalization. Combining scenarios with a unified model and framework can improve content accuracy.

“There is still a certain gap, which comes from the technological infeasibility and the lack of a good method to achieve this. This is what we need to look at now, whether new technologies can bridge this gap. We believe these issues will soon be resolved,” said Xu Qingcai.

The second issue is addressing the high cost of computing power and the scarcity of training resources.

Zhang Xin, co-founder of an AI computing power company, mentioned to TiPost Focus that for the GPT-3 model, training on an existing thousand-card cluster for a month costs over $12 million in total, with a single training cycle taking one month. In the first half of this year, the entire industry (training cards) has experienced continuous price hikes of over 25%. However, even in this situation, no one has been able to use commercial domestic chips to train large models.

Among the three elements of data, computing power, and algorithms, computing power is the foundation and competitiveness of large models. However, domestic chips are still lacking in software adaptability and stability compared to Nvidia graphics cards. Zhang Xin believes that the decoupling ability between domestic chips and Nvidia graphics cards is weak. They believe that in the coming months, they may gradually use domestic chips to train models of up to billions, or even larger scales, but accumulating computing power is still an important challenge.

Kong Dehai, co-founder and co-CEO of Lisan Technology, believes that the problem of computing power can be solved from four aspects: first, collaboration, where many calculations can be run in the cloud and coordinated according to demand; second, miniaturization of models, where small models can run on a single machine with high-quality data; third, retraining, where repeated training can help improve the user experience under limited conditions; and fourth, integrated computation.

Currently, the main computing power for AI large-scale models is in the training and inference parts. The highest cost is in the early model training, with most of it using intelligent computing centers or self-funded servers with NVIDIA A800/H800 graphics cards, or using more affordable cloud servers for training. The inference part requires less computing power and is not expensive. Most model applications require a hybrid mode of public and private clouds, and the purchase of certain cloud services to better accommodate large model applications.

Finally, there is the issue of pricing.

Pricing is the most important factor in the commercialization of large models. Due to high training costs and difficulty in data selection, the price for models with billions of parameters can be as high as tens of thousands of yuan, and the high price makes many customers hesitate to purchase.

Dai Wenyuan told TiPost App that not all scenarios or customers can accept the cost of models with billions of parameters. This is a choice that customers need to make. Even if the parameters are in the billions or trillions, it only represents their highest capability, but not all scenarios can necessarily release the technology to customers. Furthermore, the data generation scale for vertical large models will be smaller, the scenarios will be more user-friendly, and the thinking ability of Chat will be higher.

For example, Bloomberg previously released the large financial model BloombergGPT, which is applied in its vertical field; Medialink also released the first medical language model MedGPT in China, which can play practical clinical value in real medical scenarios. Large vertical models are needed in fields such as medical, finance, and e-commerce.

Several AI industry insiders told TiPost App that from an industry perspective, a general model is like an “encyclopedia,” able to answer any question and adaptable to different industry environments, while a vertical model is like an expert in a single field, which is professional but has a limited audience. However, the development of large vertical models will continue to improve the performance of models in various fields.

On June 16 of this year, OpenAI made an update, reducing the price of the GPT model by 75% and the input token price of GPT-3.5-turbo by 25%, with the latest price being 0.0001 US dollars per 1k token. Ultraman also mentioned that OpenAI is developing new technology that will allow models to be trained with less data and lower prices.

“When the model is large enough, it can generalize the problem into a common problem and output it naturally. Perhaps in the future, more than 99% of common objects or events can be handled by a single model. The benefits are that it is likely to accelerate commercialization and bring better technological capabilities. Compared with the original method, it may shorten the cycle of industrial application.” Yang Fan, co-founder of SenseTime and president of the Large Device Business Group, told TiPost App.

Zhou Hongyi, founder and chairman of 360, recently stated that the emergence of ChatGPT represents the coming of the super AI era. Large models belong to general artificial intelligence and have surpassed humans in many dimensions. At the same time, large models are industrial revolution-level productivity tools that will bring about a new industrial revolution and empower various industries. They can play an important role in the process of transforming the real economy into digital and intelligent.

“I believe that there are no insurmountable technical barriers for China to develop large models. We should thank the success of OpenAI for indicating the technology direction and route for us. Chinese technology companies have great advantages in productization, scenarization, and commercialization. I firmly believe that we can build this large model.” Zhou Hongyi stated that China will not have only one large model in the future.

However, from an investment perspective, Wei Zhe, chairman and founding partner of Jia Yu Capital, recently mentioned that “we do not touch large models.”

Wei Zhe believes that after many years in the internet industry, it has become clear that the top players always occupy 60-70% of the market share, without exception in areas such as search engines and e-commerce. The same applies to artificial intelligence, and it is difficult for winning large models to exceed two in China, and even in the world outside of China, including the United States.

The large model is a typical winner-takes-all field. It requires more money, more computing power, and more talented people. Better computing power means more people use it, and more people using it means more data. More data means better computing power results. Large models are bound to be a battleground for giants. Giants have money, technology, and most importantly, data.

Regarding the current “battle of the models,” as Zhou Hongyi said, the key to large models is to allow more people to use them, combine large model capabilities with more scenarios, and create more landing applications.

Therefore, to summarize, only a few companies can do large models, and there are few opportunities for start-ups. It can even be said that if a company cannot commercialize large models, it will definitely lose in this round of competition.

更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App

​Read More