This article was originally written in Chinese and translated using ChatGPT.
In September 2022, Shanghai’s workers rejoiced as the city’s metro finally began supporting Alipay and WeChat for QR code entry.
Previously, accessing the Shanghai metro via phone required either NFC compatibility or using the city’s self-developed app, “Metro Daduhui,” which was so notoriously user-unfriendly that it became a frequent target of complaints.
During the four years when “Metro Daduhui” was heavily used, it amassed countless negative reviews on the App Store, leaving it with a current rating of just 2.7. Similar experiences occurred in other cities, where local transport authorities resisted integrating Alipay and WeChat QR code functionality into their systems.
The reason is simple: If we view the metro as a “solution to help passengers get from point A to point B,” the “payment collection” aspect of this solution is likely only 1% of its overall importance and implementation cost. It is far too insignificant to wield influence.
A more direct reason is that local transit groups even see placing ads in their apps as a legitimate revenue stream—after spending so much money building the metro, and struggling to recoup costs, they argue, “What’s wrong with letting users glance at an ad or two when scanning codes?” Moreover, they are almost certain that even if they refuse to integrate Alipay or WeChat, city residents won’t stop using the metro. They won’t.
Now, let us bring this logic into the internet industry. I previously shared this perspective:
Any product attempting to disrupt Office that does not resemble Office proves it is doing something wrong.
If you have used computers for long enough, you have probably noticed a phenomenon similar to this argument. Historically, countless office software products have claimed to challenge Microsoft Office, but aside from WPS achieving success in China, there are few cases of substantial impact on Microsoft Office.
The reason WPS succeeded was due to a major overhaul in WPS 2005, making it fully consistent with Microsoft Office in terms of interface, functionality, and file protocols. In other words, WPS became Microsoft Office in order to disrupt Microsoft Office.
In the era of cloud documents, the number of products claiming to replace Office has increased, and while they have achieved success in niche markets, they remain far behind Microsoft Office in terms of standard-setting and market share. In fact, the only common document exchange method among these incompatible cloud document platforms is exporting to a Microsoft Office format for import on the other side, further cementing Microsoft Office’s dominance in the document market.
Why is this the case?
It is because Microsoft Office, as a 34-year-old application with super forward compatibility and ever-expanding functionality, accommodates almost all the electronic document needs of the past few decades. Its reputation for being “difficult to use” stems from the fact that each Office user typically utilizes only 10% of its features, while the remaining 90% go unused.
So, can we eliminate this 90%? No, because within the 10% of features each user needs, 3% may overlap with the 90% unused by others.
In other words, Microsoft Office represents the union of humanity’s document needs, and any challenger’s optimization of Office is essentially a subset of these needs.
On one hand, creating a product as feature-rich as Office would result in a user experience that likely would not differ much from Office (as seen with WPS). On the other hand, very few engineering teams can fully replicate the complexity of Office’s architecture—certainly not the small teams in Silicon Valley that claim to have advanced technology.
At this point, we can circle back to the title’s premise:
AI in the field of documentation is akin to Alipay or WeChat QR code entry in the metro, while the experience of the IDE document editor itself resembles the metro tracks and carriages.
AI should only serve as a button within a mature Office product, because in the overall solution of “intelligent documentation,” the key challenges and the users’ actual needs are primarily fulfilled by the documentation itself.
If AI is not user-friendly, users will simply revert to manually creating documents, as white-collar workers have done over the past 30 years. However, if the documentation process is subpar, no matter how advanced the AI functionality is, it won’t meet the users’ needs.
Next, let’s delve into a more specific example: n8n and Dify.
If you’ve been following the application of AI in personal productivity over the past two years, you’ve probably come across these two solutions on platforms like Xiaohongshu, Douyin, or Bilibili. These are two popular low-code tools that allow you to connect various online tools with AI to achieve automation, without requiring programming skills.
Among these, Dify is more prevalent in China, not only because of its Chinese localization but also because it was launched in May 2023, during the latest wave of AI advancements. Its positioning is clearer—collaborating with AI to accomplish automation tasks, making it more AI Native.
On the other hand, n8n was launched in 2019, three years before OpenAI introduced ChatGPT. Thus, n8n cannot be considered AI Native by any means.
In the “low-code automation” solution, how much does AI actually account for in the entire process? Based on my experience, if we count nodes, an “AI workflow” typically uses AI nodes only once or twice, while the remaining nodes handle linking with other tools and data processing.
In its first year since launch, Dify has developed rapidly. However, as of now, the non-AI nodes it integrates are less than half of what n8n offers. This means that even in addressing the “AI Native” need of integrating tools with AI, its upper limit does not match n8n’s. In a way, Dify is like Alipay or WeChat—it aims to improve the user’s payment experience (AI experience) by essentially rebuilding the metro system (integrating non-AI tools).
Of course, I don’t deny that Dify has many advantages. For instance, its support for AI is undoubtedly better, its interface is much more user-friendly than n8n’s, and it may continue to improve in the future. With an active community, it could surpass n8n in terms of integration. However, at least for now, it cannot achieve what n8n can, which might lead to extra effort when using it.
For example, if you want to write AI-generated content into Notion, in n8n, you just need to add a built-in Notion node. But in Dify, you would have to use an “HTTP Request” node to manually configure a request to the Notion API. (This example is valid until December 11, 2024, as Dify might update later.)
This example is not intended to disparage Dify. As an equally excellent low-code tool, Dify has already surpassed n8n in popularity on GitHub. This indicates that with community contributions and ecosystem support, it could very likely surpass n8n at some point in the future.
But as a personal user, why not wait until it truly surpasses n8n before adopting it?
Returning to my judgment of the title’s premise: if your needs align with a category where there is a mature solution and an AI Native newcomer, you should unhesitatingly choose the former, waiting for it to supplement AI functionality, rather than trying the latter and waiting for it to perfect its main functionality.
This logic can even be established at the level of business decision-making.
Anyone who has followed financial or tech news would know that this year, the AI product market in China is heavily inflated. This bubble doesn’t mean the AI technology itself is overhyped, but rather that companies are investing in user acquisition at a cost far exceeding their returns. For example, media reports indicate that in October 2024 alone, combined ad spending by several AI products reached 350 million RMB.
The result of this spending spree is as follows:
Among these, the most striking case is Tencent’s Yuanbao. As a product launched by a major corporation, it ranks even below Tiangong AI from Kunlun Tech. Of course, in line with corporate PR rhetoric, Tencent would not acknowledge such third-party data. But let’s recall—while browsing WeChat Official Accounts, Video Accounts, or Bilibili, have you ever seen Tencent Yuanbao prominently featured in the overwhelming AI advertisements?
This is because, from Tencent’s perspective, ChatBox-style AI products do not seem worth the financial burn as AI Natives. To put it plainly, even with billions spent in a cash-burning war, no ChatBox category product can achieve a WeChat-level entry point. So why not save that money and invest it upstream in models? Once Hunyuan becomes more mature, simply place a Yuanbao ChatBox into the WeChat Discovery Page. This isn’t Tencent’s first time using such a strategy—WeChat Pay and Video Accounts are, in some ways, successful examples of this approach.
However, some may question whether ChatBox products truly lack the potential to create a new user entry point. To explore this, let’s look at the situation across the ocean. I previously mentioned that OpenAI is currently facing four significant challenges:
- The model lacks a moat-like competitive edge compared to Claude, Google, and Meta;
- Rapid talent attrition;
- Consumer-facing products lack significant network effects;
- Enterprise-facing business is overshadowed by Microsoft.
The issue corresponding to this article lies in the third point.
In January 2023, two months after its launch, ChatGPT reached 100 million users, becoming the fastest-growing consumer app in internet history. By August 2024, its weekly active users exceeded 200 million. These dazzling numbers have become the basis for decision-making in AI product investment wars—if more users can be captured, the initial investment can surely be recouped later.
However, this narrative of “burning money upfront to profit later” has long been outdated in the internet industry. Specifically, ChatBox products lack the conditions to fulfill this narrative because they do not possess network effects.
Communication and social tools, once they achieve a sufficient user base, are hard to replace because when all your friends are on one platform, it becomes challenging to switch to a new tool where you cannot connect with your existing contacts.
But AI ChatBoxes lack such a “lock-in mechanism.” If any competitor of ChatGPT provides a better model or lower price, users could leave within a month. In fact, over the past year, we have witnessed Claude, Gemini, and Grox eroding ChatGPT’s consumer user growth.
Otherwise, ChatGPT’s user count would not have just reached 200 million weekly active users two years after its launch and after significantly lowering the barriers for free usage. During the same period, Google Gemini’s active users grew from 0 to 42 million, and Claude’s active users grew from 0 to 54.4 million.
To understand this issue, imagine yourself as a Western user who has been using X (formerly Twitter) and Facebook for years. You wouldn’t stop using X and Facebook simply because you started using ChatGPT, as X and Facebook are where your daily contacts reside—AI cannot replace these real social connections. However, if one day a new AI appears in X and Facebook’s messaging interfaces, you might give it a try, and if its experience is comparable, better, or cheaper than ChatGPT, you are likely to stop using ChatGPT altogether.
This appears to be the model Tencent is betting on.
After the release of ChatGPT in 2022, many optimistically believed that every product could be reshaped using AI, making the concept of AI Native an overnight sensation.
However, by 2024, practical applications of AI have proven one thing: the vast majority of AI is not a product but merely a feature.
If we accept this framework, we can see that some products in the current AI application domain clearly exhibit signs of a bubble, such as Perplexity, which is often seen as a challenger to Google. As of now, Perplexity’s search results are sourced from Bing and partially from its own index. Its contribution to AI search lies in engineering improvements on the results returned by existing search engines. This means traditional search engines are the critical upstream providers for Perplexity to deliver its product. Anyone with basic IT knowledge would understand that search engine technology is not as simple as it appears. For instance, despite Microsoft’s years of effort to catch up with Google, Bing’s core search experience still lags behind. Even without considering the technical challenges of traditional search technologies, the cloud resources Google invests in maintaining its traditional index (essentially, “to find an answer, you first need to record an answer”) and sustaining operations—including compute power, storage, and network bandwidth—likely surpass the combined cloud resource consumption of all AI applications globally. The reason AI search is effective is because it leverages these achievements rather than negating them.
Investing in a product based on traditional search engines while claiming to disrupt them is clearly illogical. The best-case scenario for Perplexity is to be acquired by Google or Bing. The worst-case scenario is being seen as a direct competitor by Bing, resulting in the termination of its search API and halting its operations. A middling scenario would be Perplexity “building its own subway,” meaning fully creating its own web indexing system to compete directly with Google and Bing. Clearly, Perplexity has chosen this route.
For a long time, however, its indexing capabilities will remain far inferior to Google and Bing, offsetting the user experience improvements brought by AI. For example, Perplexity’s executive Alexandr Yarats admitted in an interview that their current indexing range is far below Google’s, making it difficult to address long-tail queries.
Over the past few decades, the way white-collar work is performed has deeply shaped human thinking and capabilities. In other words, today’s office workers resemble assembly line workers for mental labor rather than craftsmen relying on unique skills.
Therefore, AI’s replacement of white-collar work does not require redefining the way work is done but instead quietly integrates into existing workflows through “seamless access.” In this process, rather than saying AI has disrupted the entire “abstract machine” of the workplace, it is more accurate to say that we have finally found a core gear capable of operating 24/7 without rest in the domain of mental labor. Previously, this component was the white-collar worker themselves, which meant that any attempt by businesses or management sciences to optimize the “machine” of mental labor had to consider that this “component,” being human, required rest.
This substitution does not rely on revolutionary paradigm disruption because our paradigms have long evolved to allow for the insertion of an almost tireless core component (996).
So why rebuild the machine when you can simply replace it with a 24/7 online component?
精选评论