Generative AI: A Goldmine or a Bubble?

Zhu Xiaohu, the managing partner at GSR Ventures, stands as one of the most successful venture capitalists over the past decade, having invested in and incubated numerous unicorn companies, including notable brands like Didi, Ele.me, and Xiaohongshu. His investment style is characterized by rapid decision-making and assertive action, earning him the industry moniker “Unicorn Hunter.”

In 2011, he invested in Ele.me, turning over $5 million into a return of several tens of times; in 2012, his $7 million investment in Didi yielded a return of over a thousandfold.

Recently, Zhu Xiaohu, alongside students from Heima, visited Silicon Valley in the United States, touring companies and institutions such as Nvidia, Microsoft, Meta, Google, OpenAI, Apple, and Stanford University.
The most direct realization from the trip was that the profitability direction in AI entrepreneurship has fundamentally shifted.

Upon returning, he shared his insights in June at the Heima Course, under the theme “Generative AI: Goldmine or Bubble?” The following is the edited and organized version of his insights for a detailed read, offering inspiration to all.

Since March of last year, the wave of generative AI has starkly highlighted China’s noticeable lag in AI technology compared to the United States, engendering widespread anxiety.

Amidst this anxiety, numerous entrepreneurs have recklessly invested in foundational AI technologies. Although this has led to the spectacle of the “Hundred Model Battle,” it also resulted in the squandering of societal resources. This year, many large model startups have encountered significant operational challenges.

It is evident that the direction for profitable AI entrepreneurship has shifted completely.

Currently, China’s economy is entering a new cycle, transitioning from the PC and mobile internet economy to an AI-driven economy.

This cycle presents both opportunities and challenges. Seizing the momentum of business trends and advancing with the wind is a skill every entrepreneur must master.

AIGC Startups Return to the Essence of Business

Recently, there has been a particularly evident sensation: this year marks the beginning of AIGC startups returning to the essence of business.

Without the continuous infusion of a healthy business model, standalone large model companies struggle to achieve longevity.

Independent large model companies, lacking proprietary data and user scenarios to form a data loop and continuously optimize, find it challenging to maintain lasting defensive barriers.

Among the top-tier large model companies in the United States, the leading four indeed raised substantial funds and each has secured significant partnerships. However, the second-tier large model companies, while potentially still possessing team value for acquisition now, may become valueless by next year.

Acquisitions aimed solely at extracting AI talent by major corporations are unlikely to command a high premium. For instance, Inflection, once a unicorn company into which Microsoft invested nearly 1.5 billion USD, has recently had almost its entire team poached by Microsoft. For Microsoft, with a market value of 3 trillion USD, acquiring a company only to offer the investors a return of principal plus interest is a scenario unfolding today in the United States.

In China, large model companies are engaged in a price war, offering access to nearly hundreds of thousands of Chinese characters for merely one yuan, virtually for free. It is conceivable that by next year, this service might become entirely complimentary.

For major corporations, their strategy involves leveraging low pricing to promote their computational power and cloud services. “Purchase my cloud service, and gain free access to call upon the large model’s API.” However, such pricing strategies by these major players fall below the cost line for startup large model companies. Consequently, hardly any startup large model companies dare to follow suit.

Thus, without the continuous infusion of a viable business model, standalone large model companies find it challenging to thrive in the long term.

An intriguing observation today is that, for the first time, the homegrown ventures of China’s tech giants are outperforming their acquired entities.

Typically, China’s internet giants, due to their own teams’ lack of competitiveness compared to startup companies, opt for strategic investments in startups. Although Alibaba previously invested in five large model companies, Tongyi Qianwen’s performance has surpassed all five of those investments. This development further complicates the landscape for large model startups.

Just a few days ago, Apple unveiled its own AI features. iPhone users can now use OpenAI’s ChatGPT services without the need for registration.

The underlying implication of this development is: “I may not direct users to you, but the users still remain mine. Moreover, I can switch to another large model in the backend at any moment.”

It’s rumored that Google is in discussions with Apple regarding the cost of pre-installing the Gemini large model on iPhones. After all, Google pays Apple over 10 billion USD annually to remain the default search engine on the iPhone.

Currently, Gemini is indeed slightly inferior to GPT-4. However, if GPT-5 does not launch this year, it’s almost certain that Gemini will catch up with GPT-4 by the end of the year.

At that point, how much Apple’s backend will allocate to OpenAI, Gemini, or even Apple’s own large model remains to be seen.

Therefore, when discussing the “Value Smile Curve,” on the left side we have Nvidia, which is reaping all the profits today; on the right side, there might be application-oriented companies like Microsoft and Apple; and at the bottom, possibly, are the large model companies.

AI is not a panacea

Currently, many believe generative AI to be a cure-all, assuming that integrating AI into their products will significantly boost sales. However, this is not necessarily the case.

Recently, several Chinese consumer electronics companies have incorporated AI, believing it significantly enhances the intelligence of their products. However, the actual experience often falls short of expectations. The situation is similar in the United States; although AIGC appears impressive at first glance, its practical implementation is far from straightforward.

Why is AIGC challenging to implement? Primarily, there are two issues: first, the problem of illusions; and second, the unpredictability of outcomes.

Once AI experiences illusions, errors occur, and one cannot predict when these errors will arise. Moreover, the results vary with each instance, leading to uncontrollable outcomes.

As research into illusions intensifies, it has been discovered that the core issue causing these illusions is “dirty data.” Much of the data used to train large models originates from the public internet, containing numerous inaccuracies.

Last month, Google was the subject of a humorous incident:

You meticulously prepare all the ingredients and place the pizza in the oven, eagerly anticipating a delicious dinner. However, when you can’t wait to take a bite, you discover the cheese has fallen off. Feeling disheartened, you turn to Google for advice.

Google responds, “Add some glue, mixing about 1/8 cup of glue with the sauce. Non-toxic glue will be more effective.”

This response was actually a joke made by netizens over a decade ago, but the large model considered it to be correct.

China’s AI technology is not inferior to that of the United States. What’s more critical is the data

During my recent visit to Silicon Valley, local entrepreneurs expressed skepticism about whether GPT-5 would indeed be released by the end of the year. Even if it were to launch, they questioned whether there would be a significant improvement in its reasoning capabilities compared to GPT-4.

The consensus in Silicon Valley is that it might require at least 100,000, or even 200,000 to 300,000 GPU cards to achieve a noticeable performance enhancement. However, even with sufficient hardware, the availability of data for training might still be a limiting factor.

Have you experienced “Ke Ling,” recently released by Kuaishou? The AI-generated videos by “Ke Ling” surpass those of Sora in quality.

Why does Kuaishou outperform OpenAI? Because Kuaishou specializes in short videos and possesses a wealth of data for training. In contrast, if you ask Sora if it has utilized YouTube data for training, it would hesitate to confirm.

Kuaishou has proven that having a multitude of data is more effective than merely having a vast array of hardware. With more data, I can train a superior model.

Thus, China’s prowess in AI technology is not inferior to that of the United States; more importantly, the key lies in the data.

Currently, GPT-4 can meet most needs in various textual scenarios. The greatest challenge lies in multi-step reasoning, which remains inadequate. Complex logic requiring multi-step reasoning, if each step achieves only a 90% — 95% accuracy rate, could result in an overall accuracy plummeting to below 50% after several iterations.

Chinese enterprises, by training models with proprietary data, can significantly reduce illusions and increase accuracy. The crux of the matter is whether they have accumulated enough high-quality data.

Why were companies engaged in WeChat marketing able to replace 50% of their staff with large models last year so effortlessly? It’s because they had compiled extensive data from interactions between their teams and users.

China does not lack technology but rather the creation of user experiences that evoke elation

At the Apple event, two concepts were redefined: firstly, the meaning of AI was reimagined; secondly, Apple demonstrated the principle of “entry is king.”

For most enterprises, the core focus is not AI technology but user experience.

Mastering user experience is the most challenging aspect, and it’s an area where Apple excels.

Recently, while with startup entrepreneurs in Silicon Valley, some expressed optimism about AI’s potential in consumer electronics and AI pets. Indeed, there has been a surge in enterprises combining consumer electronics with AI.

Regrettably, I feel that their offerings have not yet reached the desired level of experience. I purchased an AI pet dog, which turned out to be less intelligent than expected, essentially lagging behind an entire generation in technology.

To persuade consumers to eagerly make a purchase, one must create aspects of the product that elicit excitement and delight, a feat that is far from easy to achieve.

“Entry is king” refers to the strategic control over the primary access points through which users interact with technology or services. Despite Apple’s choice to integrate OpenAI, it does not directly funnel users to OpenAI. Moreover, I have the flexibility to switch between large models in the backend.

During our time in Silicon Valley, internal sources indicated that Apple’s backend might be utilizing ChatGPT for 70% of its operations and Gemeni for the remaining 30%. By the end of the year, if Google is willing to offer more favorable terms, the distribution could shift to 70% Gemeni and 30% ChatGPT.

For most entrepreneurs, it is imperative to concentrate on niche verticals, as the value of creating generic large-scale models has nearly diminished.

Focus on vertical applications, prioritize scenarios, and regard data as supreme

In the current landscape, generic large models are predominantly the domain of industry giants. However, China’s vertical scenarios are exceptionally diverse and data-rich, necessitating a focus on these specific contexts.

Let me share an intriguing case study of a company that maintains information systems for Chinese power plants, with a particularly interesting focus on work orders.

Power grid maintenance is inherently high-risk, prone to accidents that can have fatal consequences. Traditionally, issuing a work order for power plant maintenance, which could include up to two or three hundred steps, required at least two to three days and involved repeated manual verifications.

Last year, they spent a few months enabling a large model to learn from work orders issued over the past few years, reducing the time to issue a work order to just two minutes. Six months later, they found that the model’s accuracy surpassed human-set standards.

This exemplifies how niche verticals are inaccessible to outsiders and only by focusing on specific industries can such opportunities be discovered.

Upon identifying such a cutting-edge niche, swiftly establish a data feedback loop to secure and engage your customer base.

For entrepreneurs, finding a promising niche does not necessarily require heavy investment. If a team of ten cannot identify a lucrative scenario, expanding the team to a hundred will likely result in wasted resources. Thus, it is a misconception for companies to believe that investing millions to assemble a team of 100 or 200 people for AIGC development is the correct approach.

Consider the case of HeyGen, a company originally from Shenzhen that relocated to the United States, which specializes in creating digital humans — a seemingly simple concept. Previously, producing short videos for platforms like Douyin or TikTok involved time-consuming, repeated recordings by real people.

With HeyGen’s technology, one only needs to upload a few photos and input video script to automatically generate a one-minute short video. While digital humans have become quite affordable in China, they still command a high price in the United States. With such a straightforward product, HeyGen managed to achieve an annual revenue of 35 million US dollars in a short period.

Therefore, the focus should be on rapidly developing a Minimum Viable Product (MVP) to facilitate quick experimentation and iteration. HeyGen, for instance, iterated 30 versions within six months.

After identifying a promising scenario, the key is to test if customers are willing to pay. As long as there is a willingness to pay, you can delve deeper and further enhance your offerings.

I believe that in the current wave of generative AI, Chinese companies are not far behind their American counterparts, especially this year. If GPT-5 does not emerge by the end of the year, China is poised to lead in applications starting next year.

Let me highlight another highly successful scenario. Live commerce in China is booming, especially during major shopping events like the 618 Festival, attracting a massive influx of users. Previously, customer service personnel could not respond in time, resulting in lost opportunities.

Now, AIGC robots automatically answer questions in live broadcasts, and such products are selling exceptionally well. This scenario seems simple but addresses a significant pain point. It is advantageous for training as past product descriptions and dialog records can efficiently train vertical models to avoid hallucinations; it also tolerates minor errors well.

Incorporating AI functionalities into software for many vertical industries is straightforward once a pain point is identified, as the accumulated vertical data becomes a competitive edge. Conversely, it is challenging for AI startups to quickly identify lucrative scenarios and obtain vertical data.

AIGC will represent the long, accumulating snowfall over the next decade

Recently, Nvidia’s market value ascended to the top globally, mirroring the past two decades. At the outset of each new cycle, be it the PC internet or mobile internet era, semiconductor and hardware technologies have always experienced the most significant gains.

Back in 2000, Cisco was once the global leader in market value. However, it was swiftly surpassed by application-layer companies like Google, Apple, Facebook, and Amazon, which generated value tenfold that of their predecessors.

Furthermore, in many scenarios, domestic open-source models are now on par with proprietary models, fully capable of supporting the development of AI applications in China.

Particularly in the realm of Chinese knowledge, Alibaba’s Tongyi Qianwen outperforms Llama 3. Consequently, many startups are leveraging open-source models with 10 million parameters to train their own vertical models.

I believe that AIGC will be the extensive, accumulating ground for opportunity over the next decade, with the application layer generating the most value.

In the coming ten years, AIGC will redefine all software, consumer electronics, and end-user applications, presenting numerous opportunities within this realm.

Finally, three pieces of advice for all entrepreneurs:

1. Companies that do not embrace AI will undoubtedly be eliminated.
2. Do not blindly worship AI; focus on cutting-edge scenarios and implement them swiftly.
3. Enhance user experience and create a closed-loop data system, avoiding investments in foundational technology.

Curated by Alibaba Global Initiatives