The Dangers of a Too-Hot ChatGPT Craze
甲小姐对话张斯成：ChatGPT过热容易导致错误判断 | 甲子光年
Probably the most interesting Chinese article on the ChatGPT craze I’ve read. Except for the title, the below is translated in full by ChatGPT, with less than 10 edits. Certain lines are emphasized by yours truly. Please comment below if you see any major errors. Original Chinese from Jazzyear here.
At this moment, everyone is going crazy for ChatGPT.
Looking around the world, first Microsoft and Google, the two international giants, rallied around it, with a group of technology leaders coming out with famous quotes. Bill Gates publicly emphasized, "Artificial intelligence like ChatGPT is equally important as personal computers and the Internet." Elon Musk exclaimed "scary good" after using it and claimed that "we are not far from dangerous AI." Huang Renxun, CEO of NVIDIA, also recently stated that ChatGPT is the "iPhone moment" for the development of the artificial intelligence industry.
In China, major internet companies such as Baidu, Alibaba, NetEase, 360, and ByteDance have all expressed their intention to develop around ChatGPT. The venture capital community has also begun to appear with a variety of "Chinese versions of ChatGPT."
It seems like everyone is betting that ChatGPT can open the door to universal artificial intelligence (AGI) for humans. But the fact is, ChatGPT is still too imperfect: it cannot answer all questions and can sometimes talk nonsense; even with the powerful cloud capabilities of Microsoft Azure, it still could not escape the fate of crashing when the traffic suddenly increased. Additionally, there is the issue of money - rumors have it that OpenAI spent millions of dollars during the training cycle of GPT-3.0, but no one can accurately say how much money OpenAI has spent on the GPT model since its launch.
Has ChatGPT become too hot? When the noise blurs ChatGPT's true identity, Zhang Sicheng, the guest in this dialogue, said that "overheating can lead to erroneous judgments."
Zhang Sicheng was a co-founder of Pony.ai, the former Vice President of DingTalk at Alibaba, and the Chief Strategy Officer of Minglue. He graduated from the University of Science and Technology of China in 2001 with dual degrees in computer software and journalism, and received a PhD in pattern recognition from City University of Hong Kong in 2005.
Zhang Sicheng has always been a deep thinker. During the past three years of the pandemic, he has been living in Hong Kong, observing from the sidelines, waiting for a real opportunity. He began to pay attention to generative AI since the beginning of last year, believing that AIGC is a good opportunity. At that time, he did not expect that a "monster" would emerge six months later.
In this article, Ms. Jia talks to Zhang Sicheng about his perspective on ChatGPT and OpenAI behind it, searching for a real opportunity.
This article's outline:
Out-of-the-box thinking: "The action is so fast that it doesn't look like Microsoft, this is a winning AB strategy."
Situation Analysis: "American-style vision, Chinese-style mission."
Business Discussion: "10 Billion Dollar Club."
Technical Discussion: "Non-GPT is not enough."
Value Discussion: "It is not yet time for the internet to die of old age."
Out-of-the-box thinking: "The action is so fast that it doesn't look like Microsoft, this is a winning AB strategy."
Miss Jia: Back when Xiaobing was launched, it wasn't very popular, but this time ChatGPT has gone viral. Apart from differences in technical level, what are the other reasons?
Zhang Sicheng: First, ChatGPT has an extremely simple interactive interface. Users can open the interface and input questions to get answers. Simplification increases the possibility of going viral - all the phenomenal products on the internet in the past few years have been simple.
In comparison, Xiaobing introduced a cartoon image, which was polite and raised people's expectations. OpenAI has no "character", but instead performed beyond expectations. However, from a product perspective, ChatGPT is very ugly. Altman (OpenAI CEO) himself said it was a bad product and sometimes outputs incorrect information.
Secondly, it is very meme-able. ChatGPT can talk nonsense in a serious manner and give some interesting answers. It has topicality and is easily spread through social media. Self-media will also hype it up, and with the help of big names such as Bill Gates and Musk, it invisibly accelerates its going viral.
Thirdly, ChatGPT is very human-like. It interacts with you like a person, for example, it will tell you that it cannot answer this question, or does not know the answer, so when it makes mistakes, people will have a high tolerance.
The feelings of ordinary people towards "intelligence" and the definitions of intelligence by professionals are different. The general public's definition of intelligence is "to be like a person", that is, to be able to "express oneself like a person". Just like raising pets, we feel that pets are smart and can "understand human nature", rather than having to know that 1+1=2 to be intelligent. Now, when we look at machines, we feel the same way. This is a God's-eye view. We feel that humans are smart, so as long as pets or machines have certain "human-like" characteristics, we consider them intelligent, cute, and willing to interact with them. However, this was not intentionally designed by OpenAI, but ChatGPT was able to achieve this effect.
Finally, OpenAI's decision to launch ChatGPT at the end of last November was also a very smart choice. I believe Altman and Satya Nadella (Microsoft CEO) must have discussed when to launch it. According to reports, in fact, Altman demonstrated this product to Satya Nadella and Bill Gates last summer, so they were so confident in pushing for the $10 billion investment, probably seeing that ChatGPT and GPT would have a huge positive impact on the entire Microsoft business system and provide significant opportunities for competitive strategy.
Miss Jia: Assuming ChatGPT was launched in July last year, wouldn't it be as popular as it is now?
Zhang Sicheng: It's possible they won't release it yet. Firstly, they may not be fully prepared technologically, and recently it stopped working on Microsoft Azure cloud due to too many requests. It's estimated that the demo shown last year was using GPT-3.0, and the one released at the end of the year was GPT-3.5, which has many improvements, especially in the Chain of Thought (CoT) ability. Generally, only large models with over 100 billion parameters may exhibit powerful abilities like CoT, and the emergence of CoT may be the key to ChatGPT producing similar "reasoning" effects. Therefore, if it had been released in July last year, the interaction effect may not have been as good.
I'm not sure how they considered this time point, but they definitely designed it - after ChatGPT became popular, Microsoft's news of investing 10 billion dollars followed, and almost at "lightning speed", ChatGPT was integrated into various product lines, including Office, Teams, Dynamics, Bing, etc. These are all important products of Microsoft - the action is so fast that it doesn't look like Microsoft.
Miss Jia: Can you elaborate on why "the action is so fast that it doesn't look like Microsoft"?
Zhang Sicheng: For a mature company to complete a project, it takes more than two to three months from conception to completion. For example, Office 365 is crucial to Microsoft's revenue. Integrating external models like GPT into it is not just as simple as connecting to an API. Therefore, it is not only necessary to ensure seamless integration between the two products, but also to ensure compatibility with existing customers and services. There must be a lot of testing and debugging in the middle, and two to three months is already the limit in engineering. Therefore, I estimate that Microsoft may have secretly prepared for several months.
Miss Jia: Is Microsoft's decisive "all-in" on ChatGPT related to CEO Satya?
Zhang Sicheng: Pichai (Google CEO) and Satya are two completely different styles. Pichai is relatively conservative, but Satya is a reformer.
When Satya took office, Bill Gates asked him to make changes. He later wrote a book called "Refresh", which talks about how to turn Microsoft, an old IT company for decades, into a competitive new company.
Since taking office, he has accepted the failure of Windows Mobile, acknowledged that the acquisition of Nokia was a failed transaction, and began to actively participate in the construction of the open-source community. At the same time, in order to establish a culture of innovation internally, he publicly stated that he likes Apple's iPhone because it has all of Microsoft's apps.
His biggest achievement at Microsoft is to revive Azure. Microsoft's basic business is Windows and Office, and Azure has not been very successful at Microsoft. However, with the advent of the cloud era, the first two things are no longer effective, so he has cloudized everything and shifted everything to Azure. Now, the Windows operating system hardly makes any money, and Office 365 and Dynamics 365 on Azure are the main sources of income. Dynamics 365 is not common in the Chinese market because it mainly serves international customers, but its positioning is to help Microsoft deepen its enterprise services and gradually penetrate and occupy customers' business scenarios.
So the deep reason is still people. I think there are three people in Silicon Valley who are almost at the same level: Musk, Altman, and Satya. Musk is my idol, Altman is a person who has the opportunity to catch up with Musk, and Satya may not be at the same level as them, but he is also very powerful. He can make elephants dance and bring about such a big transformation in Microsoft, which is not easy.
Miss Jia: All three of them are extremely daring.
Zhang Sicheng: They have courage, ideas, and strategies.
Satya's strategy this time is too obvious. I am a strategist, and I don't believe that this matter has not been planned. I guess he not only planned it, but also simulated and rehearsed it many times behind the scenes, because this is a winning AB strategy --
From perspective A, Microsoft directly occupies a strategic high point on the road to AGI by controlling OpenAI and ChatGPT, challenging the long-term paradigm of search engines and fully utilizing Google's weakness in being conservative and slack in the past few years. If they can accomplish this in one fell swoop, they will break Google's long-term monopoly in the search field and allow Bing to regain a new life.
From perspective B, Microsoft directly attacks the pillar business of their opponents, pulling out the rug from under them and forcing the entire search industry to restructure their cost structure and consume together. Google's financial structure is clearly less resistant than Microsoft's, with stock prices continuously falling and the siphoning effect of capital and talent intensifying, eventually leading to a dramatic change within Google. This kind of victory can also change the balance of power between the two sides in the long-term war.
In addition, from Microsoft's perspective, they integrated GPT into Office because the GPT model is essentially a kind of AI capability output, and its effect on Office is the most obvious. After the new version of Office is released, I estimate that many online documents will be at a competitive disadvantage.
Discussing the Situation: "American Vision, Chinese Mission"
Miss Jia: For most people, how can they seize this opportunity with ChatGPT?
Mr. Zhang Sicheng: Currently, there are three paths: stay in big companies, start a business to do vertical scenarios, or join teams like Wang Huiwen's.
Miss Jia: Everyone is asking, who has the best chance to become the "Chinese OpenAI"?
Mr. Zhang Sicheng: Regarding who can become the "Chinese version of OpenAI," Li Zhifei (founder of AskCI) said a few days ago that it can be divided into seven factions (including his "self-made faction"). My personal opinion is that it is impossible to make a simple judgment: the big company team is the most likely to approach Microsoft-supported OpenAI at the commercial level, the national and quasi-national teams are the most likely to win policy support and obtain the final clearance right, and the civilian team is the most likely to break through at the innovation level, especially for teams that have been involved in LLM research and development for at least 12 months before today.
Overall, three capabilities are needed to make a "Chinese OpenAI": funds, talent, and scenarios.
Big companies are not lacking in scenarios or money, but with the talent bottleneck, it is difficult for civilian forces to rise because doing this requires a certain density of talent. Although OpenAI does not have many people, each person is a master, and only a group of masters can create an OpenAI, otherwise, in the end, they can only follow others.
A friend joked that Silicon Valley proposes "Vision," while China treats it as "Mission." This is the difference in the inertia thinking paradigm between the two sides of the ocean when facing innovation. Vision is paying for dreams, while Mission is for current livelihoods. Generally, what we usually put national efforts into developing are "do or die" fields that are strongly related to national security red lines, such as hybrid rice, two bombs and one satellite, chips, quantum computing, etc. I think artificial intelligence has not yet reached this level, although it will eventually.
Miss Jia: OpenAI will also be questioned about commercialization.
Mr. Zhang Sicheng: Altman has been asked about commercialization issues many times, and his answer is "I have no idea." If someone in China says this, the valuation may be immediately cut in half.
In 2019, Altman wrote a blog before and after joining OpenAI, expressing his political ideals and considerations on commercialization. In the article, he mentioned that he believed that the future capitalist system should be upgraded. Later, he launched a social experiment project on UBI (Universal Basic Income), which means that when robots or AI are highly intelligent and can do most of the dirty and tiring work, the social cost will be greatly reduced, and all people engaged in social production will hold a certain share. At this time, capitalism will undergo a fundamental change.
Therefore, Altman proposed the Moore's Law of Everything: in the AI era, social costs mainly have two components, the acquisition cost of AI and the acquisition cost of energy. These two costs will be greatly reduced, and when they approach zero, commercialization in society will no longer be a problem because the concept of "commercialization" may undergo a fundamental change or disappear.
Miss Jia: The fact is that OpenAI has already embarked on the path of commercialization with the help of Microsoft, and Musk even criticized them for violating the original intention of OpenAI.
Mr. Zhang Sicheng: OpenAI does indeed face practical commercialization issues, and they signed an agreement with Microsoft to solve the commercialization challenge in a dignified manner. The agreement they signed is also very innovative, as it is essentially like Microsoft renting the company, investing in OpenAI, and allowing OpenAI to help Microsoft make money. Once OpenAI reaches a certain size, they can buy back their shares.
This is a very clever move that leverages their strengths. Moreover, after the success of ChatGPT, people seem to have instantly believed in OpenAI's commercial potential. OpenAI is expected to generate $200 million in revenue this year and $1 billion in revenue next year. Now people are no longer asking Altman about commercialization issues, but rather when they can expect the release of GPT-4.
On Business: "The 10 Billion Club"
Miss Jia: "Now all the major domestic technology companies have announced their plans for large models. What do you think about their moves?"
Mr. Zhang Sicheng: "Big companies will definitely create homogenized platforms, rather than differentiated ones. Currently, there are papers and open source codes available overseas, so the big companies don't need to start from scratch. They can directly build on GPT-3.0. As for the strength of their actions, I have a few judgments:
Because they have pledged to go "All in AI" and the world's two largest search engines have already done so, Baidu must also follow suit in order to maintain its position in the same tier.
ByteDance and Tencent both deal with content. If they don't make their own large models, they may lose some businesses. Therefore, they must do it.
I think at least one of Alibaba and Huawei will do it, and it may be both. They are both capable. Alibaba's DAMO Academy is working on M6 large models, and Huawei has its own PanGu large model. However, Huawei is more low-key and often starts doing something before announcing it. They may not emphasize their large model capabilities, but instead work independently on some infrastructure, such as large computing chips. Alibaba may consider applying large model capabilities to the e-commerce field, as e-commerce is currently Alibaba's core business. In addition, from the perspective of cloud computing, Alibaba may value Huawei's approach. If Huawei considers large models as an important selling point for cloud resources, Alibaba will have to follow suit.
Miss Jia: "Many start-up companies are also making their moves. Can you draw a line and say how much capital these companies need to become China's OpenAI?"
Mr. Zhang Sicheng: "USD 10 billion. It's simple. Musk set up OpenAI with USD 1 billion, and then Microsoft invested another USD 1 billion in 2019. They have spent USD 2 billion, and Microsoft has announced that it will invest another USD 10 billion in the future. This is enough for OpenAI to burn for ten years.
However, I think that building large models from scratch in China may not require as much money, but it will likely still need an investment of at least RMB 20-30 billion. Although the cost of Chinese labor and electricity may be lower, the cost of training, machinery, chips, and hiring scientists is also not low. Therefore, this is a 100 billion club.
Miss Jia: "Apart from money, what other barriers do start-up companies face in building large models?"
Zhang Sicheng: "The algorithm of large models itself is not a barrier, as there are many published papers available. The barriers that exist, apart from funding, are mainly talent and data, especially experienced engineering talent and high-quality training data. Even if we fine-tune the foundation model to make vertical models, it still requires a considerable amount of talent, computing power, and data reserves."
Miss Jia: "When did you start paying attention to ChatGPT?"
Zhang Sicheng: "I saw the Gartner report on generative AI early last year. As someone in the field of artificial intelligence with a doctoral degree in a related field, I thought AIGC could bring about earth-shaking changes. It will greatly enrich content, but I didn't have time to invest at that time.
At that time, most companies in China were still following the Bert route, and only OpenAI was persistently working on GPT. When ChatGPT appeared in November, I began conducting crazy tests in various groups."
Miss Jia: "What are your plans?"
Zhang Sicheng: "I decided to step down in December last year. I couldn't miss the biggest opportunity of the next decade. But I don't think I have enough influence to build a team from scratch to develop large models, and I can't do too general scenarios. So I plan to work on some vertical scenarios. At this level, I have a lot of respect for Wang Huiwen, who stepped forward with money and people to create the Chinese version of OpenAI.
I expect to have about 6-12 months to make quick money in some 2C areas, such as projects similar to Copy.ai and Jasper.ai. There are already twenty or thirty such projects in the European and American markets, and there may be even more in China. After twelve months, when the big companies react and the team and business take shape, it will be difficult for these companies to survive. Domestic big companies also don't like to acquire such companies, so there's basically no way out. I think there are still considerable opportunities in many vertical areas, such as 2B."
Miss Jia: "Do you have a clear picture of what a technical partner should look like?"
Zhang Sicheng: "The best technical partner should be NLP-related, have experience in large models, and not necessarily be a big name. It's important to have engineering experience and stay close to the forefront to maintain sufficient hands-on ability."
Miss Jia: "What vertical scenarios do you currently see as promising?"
Zhang Sicheng: "I'm more optimistic about several categories:
The first is the education and training industry, of course, in policy-supported areas such as vocational education. Education and training can be carried out using bots (robots) to let users choose the services they need. Large language models currently show a very interesting feature, which can help reinforce cognitive education. But how to develop this field may require some skills.
The second is the design field, especially in the direction of high-dimensional design, where complexity needs to be reduced. For example, the architecture field needs various design views and rendering effects, the game field needs quick modeling and intelligent interaction, and the industrial manufacturing field needs to stimulate creativity as much as possible under conditional constraints. These can all be helped by generative AI technology to improve efficiency.
The third area is the video sector, which is the closest to making money in AIGC entrepreneurship because the business model is already quite mature. For example, for big V on platforms like YouTube or Bilibili with millions of followers, content production is a pipeline process, with more exposure leading to more income. Therefore, their demand for frequent content iteration is strong, and generating AI can be used to quickly generate ideas, collect materials, and edit drafts. Of course, there is bad news in this field, as there are giants, but the good news is that there is not only one giant.
However, I think that after this wave of enthusiasm subsides, we will see more clearly, because I also have cognitive blind spots. It's too hot now, and overheating can easily lead to some erroneous judgments.
For example, I personally believe that generative AI has great potential in the medical field, but few people discuss this area. There are two major scenarios in the medical field: medical imaging and biopharmaceuticals.
Google has already started in the field of biopharmaceuticals, and their Alphafold can predict a large number of protein 3D structures. In recent years, companies in Silicon Valley have been trying this direction, using AI to help find new drug targets or synthesize new small molecule drugs. However, this requires deep cooperation with pharmaceutical companies, as sufficient data sources are needed to train the models.
In addition, there is also great potential in medical imaging, and diagnosing images is an important scenario, such as Lianyi Medical in China, which already uses AI to help doctors identify lesions in CT images. With the expansion of large language models in the multimodal field, I believe this can become a new blue ocean.
Ms. Jia: Technically speaking, is there an optimal path among large models such as GPT and Bert?
Zhang Sicheng: In the people I've talked to, this question can be divided into three categories:
The first category is professionals, who don't think that one model is better than the other. Objectively speaking, Bert and GPT are equally strong technically, but they have different strengths. Google is working on models other than Bert, and they do acknowledge that some of the characteristics exhibited by GPT are worth studying. Currently, people generally believe that achieving AGI with large models is a very promising path, but it's not that GPT is the only option. It can only be said that, in the evolution towards AGI, OpenAI has shown with ChatGPT and InstructGPT that GPT might be a more ideal large language model.
The second category is the laypeople who may not know what Bert is, but only know about ChatGPT and think it represents the future of artificial intelligence.
The third category falls in between, which is us, people who are not that specialized but not completely ignorant. We tend to believe in GPT. I have been wanting to write an article titled "Seeing is Believing" for us pragmatists. If we see it, we believe it. Because ChatGPT's performance is obviously better than Google's, and Google has even failed, people tend to acknowledge GPT more and are more willing to invest time in studying it.
Ms. Jia: Why do people who want to build large models now all say "we lack people"?
Zhang Sicheng: Perhaps because in the past few years, 90% of companies in China that work on generative AI have been using the Bert approach. At the time, GPT's practicality was relatively low, and it was difficult to meet the needs of project-type requirements in terms of cost-effectiveness. Perhaps the entrepreneurial environment in China has always forced people to consider short-term practicality.
Ms. Jia: I heard that Huang Wei from iFlytek seriously compared GPT and Bert, but chose Bert.
Zhang Sicheng: But it is said that they started to switch to the GPT approach after seeing InstructGPT, and they have acted quite quickly in the industry. Now, there is a hypothesis of using SparseGPT for compression, which can significantly reduce training costs and the need for computing power while maintaining accuracy.
In addition, it is said that some scholars are starting from the infrastructure and proposing not to use GPUs, such as Nvidia's A100 or H100, but to use conventional computing power in cloud computing for training. This theoretically seems to have some potential, but it can only wait for the major cloud platform companies to implement it.
Value Discussion: "It is not yet time for the internet to die of old age."
Ms. Jia: What is the social value of ChatGPT?
Zhang Sicheng: ChatGPT is a large-scale social experiment. This is OpenAI's stress test on the GPT model, allowing the whole society to use it to test its universality, so that it can meet Microsoft's commitment that the product has reached commercial performance.
I believe AGI will definitely bring about a profound revolution, no less than the Industrial Revolution, but ChatGPT hasn't reached that level yet. If we want to make an analogy, I think ChatGPT's impact is similar to the birth of the iPhone. The biggest change brought by the iPhone was to change the input from buttons to touch screens, which was a revolution in human-computer interaction and led many people to think about the changes in their work and life, thereby ushering in the era of mobile Internet. This is a self-stimulating reaction that permeates every corner of society.
I think it's not the end of the internet yet. ChatGPT may bring a new generation of the internet, which I personally call the "Intelligent Internet". The most fundamental characteristic of this generation of the internet is that bots will become the most basic interface between humans and machines, no longer apps or browsers.
This change will trigger paradigm shifts in many scenarios, and even affect the classic architecture of the internet.
The explosion of ChatGPT is actually a universal social and mental education, which may have threefold effects: First, AI is a basic skill, and people need to consider how to better use it to improve their overall abilities. Second, the improvement of AI capabilities may reconstruct the paradigms of our daily life, and we need new ways to solve the problems of original scenarios. Third, it may bring more micro-innovations, making some scenarios from imagination into reality.
For our generation, "the ones who take away your jobs are not AI, but other people who master AI tools." For the next generation, "the ones who differentiate you and AI are not humanity, but the yearning for the vast universe."
Ms. Jia: Now many people are paying attention to the entrepreneurs of this wave, and the consensus is that startup companies may need to do more suitable application scenarios and vertical markets. But if a startup company doesn't do big models, how can it build barriers?
Zhang Sicheng: If it is purely a 2C business, there is no barrier, it just needs to be fast.
When I was "entrepreneurial" in DingTalk before, we believed in a saying called "the world's martial arts are fast and not broken", which means we need to run fast, not only faster than our competitors, but also faster than the platform. When I summed up this experience, I was very grateful for WeChat team's "mercy", because the scissors difference allowed DingTalk to survive.
Now the environment has changed. 2C startups have only two options when they have reached a certain level: either be crushed by others or be acquired by others. If you ask if there is still a chance to create another Didi now, I think there is no chance. In the current platform pattern in China, it is almost impossible for C-end scenarios to run independently. Because the giants are all behind, they are basically waiting and watching, and once the PMF of a certain field is verified, they can quickly crush the startup company with their resources and users.
Ms. Jia: What was the core variable that allowed Didi to emerge at that time but now has no chance?
Zhang Sicheng: When Didi started, even the giants within the industry were uncertain about whether or not to enter the market, and at that time there were many business opportunities available that they couldn't even spare resources to invest in the ride-hailing industry. But now, the giants are generally in a state of scarcity and hunger for opportunities, and once they start competing with each other, startups have almost no chance of winning.
Miss Jia: If there is a startup company that still wants to target consumers (2C), what are its opportunities?
Zhang Sicheng: If they want to target consumers, they should target the international market. If they want to do it in China, they should focus on the enterprise (2B) market.
Miss Jia: But the international market is also facing fierce competition among giants.
Zhang Sicheng: But the international market is big enough, and the giants are generally more tolerant. If a startup does well, they may just buy it instead of crushing it with their own resources, which is different from the startup ecosystem in China.
Therefore, startups targeting the 2C market can only rely on speed to quickly launch popular products and build a loyal user base to form a moat. Otherwise, they will likely be crushed by the giants in just a few seconds. However, there are still many opportunities in the 2B market. Looking at the development of cloud computing/SaaS in the past decade, even large platforms find it difficult to dominate the enterprise service market through scale, because it is a market that requires differentiated payment.
Miss Jia: What are some of the unresolved issues facing ChatGPT in China right now, in your opinion?
Zhang Sicheng: The main question that everyone is focusing on right now is how ChatGPT can be implemented in China. Because there is currently no true ChatGPT in China, Baidu's Wenxin Yiyu may face some challenges.
Another major unknown factor is how the government will regulate it.
Products like ChatGPT are cognitive educational tools, and have traces of values everywhere. In a sense, OpenAI's ChatGPT, like Hollywood's movie factories, plays a similar role, intentionally or unintentionally spreading Western values. Because the distribution of the corpus used in the training model largely determines the bias of its output.
So far, the mechanism of LLMs (large language models) is still a black box, and we cannot ensure that its output meets the requirements of cognitive education through parameter tuning. Under this uncertainty, combined with consideration for social data security, I estimate that in the future, China and the West will likely "treat the ocean as one", each supporting ChatGPT-like products.
Miss Jia: Are there any common misconceptions in the market's understanding of ChatGPT at this moment?
Zhang Sicheng: It is difficult to judge what is right or wrong at the moment, but there are two obvious misconceptions:
Firstly, there is a cognitive misconception that generative AI, represented by ChatGPT, is actually a productive tool that can help humans improve efficiency, but it is difficult to replace human roles. It may gradually replace individual types of work, and when combined with physical robots, it can help humans get rid of some work that is not suitable for humans. But for most people, it is a highly efficient tool that can improve work and communication efficiency. Therefore, the public and the media do not need to worry too much, and can try to recall the overall changes and progress of society after the launch of the iPhone.
Secondly, I think some views are too extreme. After ChatGPT became popular, this type of AI technology will indeed change many production relationships and continue to trigger a series of scenario upgrades, but it does not have the ability to significantly improve productivity, and it is far from the shock brought by the "Industrial Revolution". Therefore, there is no need to deliberately hype it.
Humans are still far from achieving AGI (artificial general intelligence), and ChatGPT only vaguely shines a light for humans moving forward in the darkness. The past decade has proven that many innovations cannot withstand hype. The more enthusiastic they are at the beginning, the more disappointing the outcome may be. For example, blockchain, cryptocurrency, and Web3.0 are almost like this. If they had been allowed to develop quietly at the beginning, they might have shown their true value.
I personally believe that the era of AGI will definitely come, but it may take 50 or even 100 years, and whether our generation can see it is uncertain. So, at this stage, we should not worship ChatGPT. Worshiping something often brings fear, which is not conducive to the development of innovation. We need to have faith, and we need a soil for our faith to strive towards.
Miss Jia: I have always believed that trial and error are two different things. Technological innovation itself requires trial and error, such as OpenAI's large-scale social experiment during the window of opportunity. But if we hold it down during the trial and error stage, treat trial and error as mistakes, and cannot collect social data and conduct social experiments, our gap may become wider and wider.
Zhang Sicheng: You are absolutely right. To some extent, premature and stringent regulation may stifle innovation. We can express optimism, but we should not worship it or demonize it. ChatGPT is ultimately still in the AI category. How it creates higher productivity, helps the country develop the economy, and helps people improve their quality of life are the key issues.