Insights From QbitAI | Public Account QbitAI
In the past three months, the wave of AIGC and the large models behind it have swept across the Chinese internet, allowing everyone to enjoy the conveniences brought by advancements in artificial intelligence technology.
At the company level, leading enterprises with large model technology and resources have gained the most attention, whilestartups in the large model application layerare active but have low visibility.
Furthermore, with the release of plugins related to ChatGPT, application-layer companies not directly involved in the competition for large models could be “overwhelmed” by a plugin from companies like OpenAI.
There is no doubt that the development of general artificial intelligence will continue to affect the survival space of application-layer enterprises.
So, how will application-layer startups find their own innovative space and build unique business models under this pressure? How will their entrepreneurial value be recognized by society and the market?
In a discussion on how application-layer startups can link large models with customers to create value, Chen Xiyan, co-founder and CTO of Yilan Technology, shared his industry experience and insights during the “QbitAI Insights” live broadcast.
The following is a summary based on the shared content:
The emergence of large models has leveled the AI capabilities of many companies, revolutionizing many AI service providers, and it is spreading at an unprecedented speed, being called the fastest-growing application with over 100 million monthly active users, disrupting many technologies, businesses, and services.
After ChatGPT became free, some startups based on the OpenAI ecosystem were indeed bewildered.Although OpenAI’s CEO stated that “we will not compete with customers,” Jesper’s CEO also said, “The emergence of more large models has allowed companies to break free from a single reliance on OpenAI,”but the logic of these explanations is actually quite weak.
The brilliance of ChatGPT lies not only in the large model itself but also in the natural language interaction model, which is the most advanced human-computer interaction mode since command lines and graphical UIs. GPT-4 provides the potential for general artificial intelligence, and the human-computer interaction in dialogue with AGI seems to become omnipotent. Domestic large models like Wenxin Yiyan and Tongyi Qianwen are similar.
Over the past six months, our team has often discussed a question: now that everyone is using large models, what exactly is our value?
Opportunities for Startups Under the Shadow of Large Models
During the Wenxin Yiyan launch conference, Li Yanhong proposed a four-layer model: chip, framework, model, and application. However, he only mentioned these four layers without clarifying whether the relationship between them is pyramid-shaped, inverted pyramid-shaped, olive-shaped, diamond-shaped, or some other structure.
Chips are undoubtedly the core competitiveness that can be a bottleneck. The framework is a game for large companies. As for the model, it currently seems to start at 1 billion, and Baidu has stated that Wenxin has simultaneously developed large models for 10 industries.
At the application layer, it currently appears that interactive applications like ChatGPT, Wenxin Yiyan, and other large models are ToC (business-to-consumer) and free. Everyone is talking about building an AI foundation and forming their own application ecosystem.
If the AI foundation is at the bottom (on the ground) in the model hierarchy, then it serves as a platform for application-layer entrepreneurs. But if the AI foundation is at the top (in the sky), what application-layer entrepreneurs will be left with is merely a shadow.
At present, it seems that the AI foundations that are building large models are exhibiting a “I want it all” attitude: claiming to be a large foundation, but in reality including all four layers of the model.
Only children make choices; having spent so much money on large models, of course, they want it all.
Shadow 1: General Artificial Intelligence Causes Many SaaS to Lose Barriers Overnight
General artificial intelligence seems capable of doing everything, leaving many SaaS software bewildered; years of accumulation may become worthless in front of large models. All software companies face a situation: being forced to embrace AI.
When it was proposed that “all applications are worth redoing with AI,” large models first integrated Copilot into all products of their parent companies, such as browsers, search engines, instant messaging, office software, and so on. This truly raises doubts about what value we still have in other software.
Shadow 2: Not Only Replacing Machines and Software but Also Quickly Replacing Human Jobs
I believe that jobs that can be simply and accurately described by input, process, and output are easier to replace.
For example, changing all blue colors in an image to red, merely doing layout editing, responding to standard documents in customer service, or even making a cup of coffee when the method of adding beans, grinding, and the amount of water and milk are all very standardized, the taste may not differ much from hand-made coffee.
Shadow 3: All the Spotlight is on Large Model Companies; Without Money, You Can’t Get Onboard
Currently, almost all the spotlight is on those companies producing large models, including startup stars, BAT (Baidu, Alibaba, Tencent) companies, and prestigious universities. One thing is certain: startups cannot participate in the development of large models.
A100s are hard to buy, but there are actually many ways to purchase them; for example, they are available on JD.com—only the last time I checked, it was 100,000 yuan each, and today it has already turned into 119,999 yuan. Whether you can buy it is not the issue; the issue is whether you can afford it.
Even if a startup has 50 million, after buying 500 units, they wouldn’t even be able to afford a network cable, and they could only use a credit service for electricity bills. Therefore, if a startup wants to enter the large model arena, they need at least a billion to even sit at the table.
AIGC has not created new demand. Asking for an AIGC painting may simply mean wanting a cheaper painting, as determined by the supply-demand market. AIGC is merely a tool, and the demand side always wants cheaper products.
AIGC has repeated many business processes and supply-demand relationships; almost all enterprises hope that AIGC can enhance productivity, but in reality, there are many difficulties.
The Real Difficulties in Productivity Enhancement with AIGC
1. Workflow Solidification is Hard to Change.
AIGC tools emphasize individuals; individuals are workers who sell time, and if they accumulate enough time, they may receive a salary. However, as a new thing, AI tools have many uncertainties, making it hard to persuade workers to increase their learning costs for an uncertain tool.
2. Business is Much More Important than Systems.
In many ToB (business-to-business) operations, once matured, business relationships often become more important than the business itself. Systems are weakened, and some bosses even require the tech department to ensure only that it does not crash, regardless of how it affects the company’s profitability. I heard the most unarguable statement: if the client’s working method does not change, what difference does it make if our company changes alone?
3. Many Workers Prefer to Use Tools Secretly.
We have to admit that in many scenarios, especially in creative work, using AI is not a very respectable thing. If workers create something with AI or use AIGC to save more time to rest or goof off, they naturally do not want to admit it and can never admit it.
4. Concerns About Information Leakage.
Security is indeed a problem, but more often it may just be an excuse; it is not as important as the previous three points.
Currently, AIGC tools empower individuals more than enterprises. I hope everyone remembers this statement, as it presents an opportunity for our entrepreneurship.
What Kind of Positions are Hard to be Replaced by AI?
1. Require Stable and Precise Results
Whether GPT-4 can solve math problems is a highly controversial issue.
For instance, if a computer can only perform the operations “+1” and “X2,” and if it is given the input “1,” asking how many steps it takes to get to 63. GPT-4 will definitely use the reverse method, and while the method is correct, it makes mistakes every time. Ironically, it knows it made a mistake and still tries to pull it back. For example, in one step, it might “X2” and get 70, then suddenly realize it exceeded the target, and at that moment, it tells you to subtract 7 to arrive at 63.
I do not believe that stability and precision are necessary attributes for a large model, nor do I believe that language models can solve math problems. This is my somewhat biased worldview, and it may be wrong.
2. Strong IP Attribute Business
Have you noticed that calligraphy, as an important art form in our country, has yet to see an AI model emerge? This may be due to value issues; a calligraphy AI model may simply lack value.
For instance, if a character is written by Qi Gong, and another by Su Shi, if a third character is written by AIGC, it is very likely to be discarded. This is because calligraphy emphasizes personal style, aesthetics, and personal IP.
3. High Security Requirements
Imagine a profession: AI hairdresser.
A hairdresser’s profession requires extreme trust and safety. If a malicious hairdresser approaches me with a knife, I might say, “Brother, don’t hurt me,” and appeal to emotions, which might give me a slim chance of survival. But if the person in front of me is a robot, merely a killing machine, I have no way to defend myself.
Moreover, the hairdressing industry also has a strong IP attribute; every hairdresser is a strong IP in the community.
4. Lack of Data Sets or Difficulty in Annotation
The generalization ability of large models is still limited by data sets. If an industry is likely to be replaced, a simple criterion is whether that industry can organize enough high-quality annotated data; if the data quality is very high, it is likely that machines will replace humans.
However, if it is difficult to annotate, such as an artwork that contains not only the creator’s emotions and feelings but also the viewer’s emotions at the moment, it is unlikely to be replaced. As we often say, “there are a thousand Hamlets in a thousand people’s hearts,” this complexity is hard to annotate effectively, and the significance of annotation is minimal.
5. Requires Hands-on Experimentation
There is a classic question: if ChatGPT had been invented 500 years ago, would humanity still know that the Earth is round?
I believe that in the face of science, AIGC and the Bible are the same; not because they are authoritative, but because they both need to be discarded in the face of science.
While AI can handle and simulate some established tasks, it lacks human intuition, emotions, and understanding of abstract concepts, which cannot be simulated.
Opportunities for Tech Startups Under the Background of Large Models
1. Understanding Business and Technology: Linking Large Models with Users
We have a client who is a listed company in the film and television industry. Initially, the boss required the entire company to use AI, and over 100 people responded by using a variety of tools, which later narrowed down to only ChatGPT and Midjourney, and eventually, they ended up using nothing. It was only later that they found us to provide enterprise-level AI tools that seamlessly integrated with their existing workflows.
Thus, enterprises will always need information systems. If someone understands both business and technology, becoming a connector between large models and enterprises is a good choice.
2. Wealth and Capability: Joining Large Model Startups
There are still many worthwhile endeavors, such as multimodal, latent information, brain-computer interaction, etc., but they require wealth and capability.
3. Interesting Souls: Using AI to Create Engaging Content
AIGC represents a particularly good era for content creators. Previously, large creative organizations may now only require a few people.
Additionally, ChatGPT’s serious nonsense or whimsical ideas are particularly suitable for content creation.
4. Straightforward Programmers: Fine-tuning/Customizing Models
There is a purely technical direction: providing fine-tuning of large models for enterprises or offering industry-customized models.
When I first saw MaaS, or Model as a Service, I felt despair, as if I were being pushed to the brink by large companies. However, recently I came across the term Fine Tuning as a Service, which I found particularly interesting and clearly conceived by programmers. This term made me feel the optimistic and positive spirit of programmers.
Which Positions Will Be in High Demand Under the AIGC Background?
In fact, the most profitable aspect of the gold rush is selling jeans. In the entire AIGC context, in addition to the direct entrepreneurial opportunities mentioned earlier, I believe there will also be some highly sought-after job positions or services that can be formed.
1. AI Alignment Engineer
As AI tools continue to develop, how AI aligns with human values and how AI systems can ensure their interpretability will become key issues. AI alignment engineers are responsible for ensuring that AI systems can always align with human values, protect human interests, and guarantee human control over AI, even to the point of pulling the plug or cutting power when necessary.
2. Model Training Engineer
As mentioned earlier, every enterprise deserves its own model; thus, obtaining a proprietary model for a company based on large models through fine-tuning or distillation is also very important.
3. Data Governance Engineer
Although enterprises deserve their own models, models rely on data.
Data governance refers to the reasonable management of data quality, data integration, data privacy, business processes, etc., within an organization, ensuring the accuracy, consistency, and usability of data. Data must be usable to have a chance to train models.
Most enterprises are insufficiently digitized. In fact, the more online a company’s business is or the more data services it provides, the lower its degree of digitization; many internet companies are even less digitized than some factories.
4. Data Annotators
High-quality annotated data will become the core competitiveness of AI systems. Recently, I saw in the news that Beijing proposed a concept to create a unified data annotation service, and it has become evident that data annotation is a very important business.
Insights from babyAGI and autoGPT
Recently, babyAGI and autoGPT became quite popular, rising quickly and cooling down just as fast. However, they left behind several valuable terms, such as “chain of thought” and “task queue,” which some places interpret as “workflow.”
The left image is the workflow diagram of babyAGI version 2, where the core is to form a list of questions from the first question, place them in a queue, execute each question accordingly, and then interpret the corresponding answers through ChatGPT, generating several questions from the answers, similar to a pyramid structure, gradually uncovering the essence of the solution.
The right image refers to the most important steps in the first version of autoGPT, which also breaks a question into multiple steps and lists corresponding plans. However, it adds an additional step of checking whether something is correct, followed by criticism and reflection. This step is crucial for an intelligent AI system.
From these two open-source projects, we can learn many aspects that large models have yet to touch upon or unique AI systems. My idea is to improve them and use expert systems to set up our business workflows, pre-entering them into the system.
The Confidence and Advantages of Creating AI Products at the Application Layer
Why does Yilan Technology have the capability to develop AI video tools at the application layer, and what advantages do we possess in this field?
The first advantage is data accumulation. Yilan has been in the commercial video business for many years, serving thousands of corporate users and signing contracts with hundreds of thousands of video creators globally, accumulating a lot of unique commercial video data that continues to grow during the creation and trading process.
The second advantage is a clear scene. We have always focused on video creation.
The third advantage is that our business is relatively mature; having been established for five years, we have accumulated a vast number of creators and content consumers.
The fourth advantage is the safety of creation. General artificial intelligence AGI has many risks and uncertainties, but the inclusiveness of content creation and the measures for AI alignment and content safety are relatively well-established, with warning mechanisms for political, violent, and pornographic content already in place.
We have chosen a suitable vertical field for ourselves and possess certain advantages in this area, which is the characteristic of Yilan’s AI video tools at the application layer.
Having said so much, let’s actually introduce Yilan’s AIGC workflow product “AI Screenwriter.”
We pre-inject the workflow into the AI system, and then the AI executes it step by step. In my opinion, the “Step by Step” process should not be executed by general artificial intelligence but should be input by human experts.
The first step is to generate several ideas from a single sentence. This idea generation process is akin to brainstorming, which is particularly suitable for AIGC. We can input several keywords like palace, melodrama, love, ethics, counterattack, twist, and it can instantly output many ideas.
The second step involves starting character design within the workflow. Character design requires certain elements, such as personality traits, behavioral habits, physical characteristics, dialogue style, including gender and age.
Now that we have characters and ideas, we can input this content into the AI to generate plots, which must include conflict and twists, as this is also something we need to pre-set. Then we can quickly generate several plots and even batch-generate scripts of several hundred or even thousands of words. We can choose the most challenging plot, and based on that plot, we can also generate a series of concept images.
Moreover, as you may know, due to token limitations, we can also decompose a plot into several paragraphs and expand each paragraph, which is also a process that can gradually enrich the script.
The four core elements of Yilan’s “AI Screenwriter” workflow are: context storage, massive creativity, visualized creation, and evaluation feedback.
Evaluation feedback can be understood as script scoring, akin to the criticism stage in autoGPT. In this process, we can generate more information and data.
After script evaluation, these plot segments undergo constant adjustments and optimizations, then proceed to the next step, once again generating visualized creations, continually refining the quality of the work.
AIGC Accelerates the Arrival of the Mediocre Era
AIGC can raise the average level of all production work or creation, but it can only make the future better than the past. However, if everyone becomes equally good, what was once better becomes mediocre now. When AI tools become fully popular, mediocre creations will become exceedingly cheap, and humanity’s future belongs to true craftsmen.
Undoubtedly, the proliferation of AI tools and the explosion of AIGC content are part of a historical trend that is irreversible and unavoidable. From media to fashion to architecture and beyond, we will be filled with inevitable and pervasive mediocrity. However, optimistic individuals will always find their own ways to create unique waves. Finally, I would like to share a recent tweet from Musk with everyone.
Maintain your craftsmanship spirit, even if you have to walk alone. Stay on the right path, even if you have to walk alone.
About “QbitAI Insights”
The CEO/CTO series sharing event initiated by QbitAI periodically invites CEOs/CTOs of cutting-edge technology startups to share their latest strategies, technologies, and products, and to discuss cutting-edge technical theories and industrial practices with practitioners and enthusiasts. We welcome everyone to pay more attention ～
If you want to watch the live replay, please click “Read the original text”~
— The End —
Click here👇 to follow me, and remember to star it~
One-click three connections: “Share”, “Like”, and “View”
Cutting-edge technological developments are seen every day ~
