Summary of Key Points
Through a month of firsthand observations in Silicon Valley, the author compares the progress of AI in March with the current state and makes several key judgments: The arrival of AGI (Artificial General Intelligence) is much faster than expected, with the rate of AI replacing knowledge workers and researchers significantly increasing; large models will consume over 90% of application value, but the remaining 10% of applications with leverage can give rise to major companies; the gap in capabilities among coding models is narrowing (with both China and the US catching up in terms of open-source and closed-source technologies); the demand for enterprise AI represents a vast yet unexplored potential, requiring professionals with multidisciplinary skills to implement it effectively; model stratification and token re-pricing represent the next significant opportunities, with great potential for Chinese open-source models. The author also notes the shift in priority within the hardware industry, where storage is considered more important than CPUs and optical components, as well as the contrast between the fervent belief in AI in Silicon Valley and a more rational approach in China.
1. AGI is Coming Faster Than You Think: The Replacement Rate of White-Collar Workers and Researchers is Soaring
In March, the author estimated that 50% of white-collar workers and 30% of researchers would be replaced by AI. Now these figures have risen to 80% and 50%, respectively, and even the automation of infrastructure tasks (such as server management) has begun (10-20%). In simple terms, while AI might previously help with tasks like writing copywriting or researching information, it can now handle half of the data analysis and initial draft work done by researchers. Nearly 80% of the daily tasks of white-collar workers (such as writing reports, sending emails, and creating PPTs) can be automated by AI.
More importantly, industry experts believe that the rule of “the larger the model, the better the outcome” (scaling law) has not yet reached its limit—there is still plenty of data available for optimization, and both algorithms and hardware can be improved, indicating that AI capabilities will continue to surge. However, this progress is largely limited to a few large companies (such as OpenAI and Google), and most people are not likely to benefit from it, much like the general public cannot participate in rocket research.
2. Large Models Are Like “Predators”: They Consume 90% of Applications, but the Remaining 10% Can Lead to Success
The author uses Liu Cixin’s novel *The Devourer* as a metaphor: Large models represent a supercivilization that consumes everything; most applications (90%) will be replaced by them (e.g., common chat and search functions, which can be directly handled by large models). However, the remaining 10% of applications with unique data or specific use cases can escape this fate through “acceleration”—the key lies in the value of that data (e.g., exclusive medical case data or financial transaction data).
For example, an application that simply writes essays using AI will likely be taken over by ChatGPT, but one that analyzes rare disease data using unique information can outperform large models and even become a successful company.
3. The Gap in Coding Model Capabilities is Narrowing: China and the US are Catching Up
The gap in capabilities among coding models (AI systems used for writing code) is closing: firstly, the difference between leading closed-source models (such as GPT-4 and Claude) is diminishing; secondly, the gap between open-source models (such as DeepSeek from China and Llama from the US) and closed-source models is also narrowing (reflecting the advancement of AI in both countries).
However, switching to new coding tools presents challenges for different types of users: individual developers or small teams can easily make the switch (since the cost of trying new tools is low), but large companies face difficulties. On one hand, costs are a concern (they need to pay again for new tools); on the other hand, employees fear losing their jobs and are reluctant to learn new skills. Additionally, once companies get used to AI systems that can automate entire work processes (from requirement generation to code deployment), it becomes even harder to switch.
4. Enterprise AI Demand is a “Black Box”: Vast Potential but Difficult to Implement
The demand for AI in enterprises does not grow gradually but in leaps and bounds. If funding comes from the IT department (for cost reduction), growth is linear; if it comes from business departments (to generate revenue, such as through optimized marketing), growth can be explosive.
The main obstacle to implementation is not a lack of model capabilities but the gap between theoretical knowledge and practical application—companies need professionals who understand AI, business processes, and management. For instance, an AI model that can write code must be integrated with existing systems and employee workflows.
There are three ways for companies to deploy AI: fully in the cloud (using cloud-based LLMs), hybrid (local hardware + cloud-based LLMs), or entirely on-premises (own hardware + open-source LLMs). Similar to the SaaS revolution 10 years ago, it is still unclear which approach will dominate, but all are growing rapidly.
5. Model Stratification and Token Re-Pricing: The Next Big Opportunities
The current issue is that the same model (e.g., GPT-4) is used for both simple tasks (like checking the weather) and complex tasks (like writing papers), which is inefficient. “Model routing” involves finding the most suitable model for each task—using a cheaper model for simple tasks and a more expensive one for complex ones, which can be 100 times more efficient than engaging in a price war.
Furthermore, tokens (the basic units of data processed by AI) are now priced equally, regardless of the complexity of the task. In the future, token pricing will adjust based on task complexity, representing a significant opportunity. For China, model stratification could open up market space for open-source models, as companies may not be able to afford closed-source alternatives. Therefore, the author is “very optimistic about the prospects of AI in China.”
Finally, the author mentions that the hardware industry in Silicon Valley prioritizes storage over CPUs and optical components because storage is crucial for AI to retain knowledge. Since memory companies operate on a end-to-end basis (without third-party manufacturing), they are likely to be more highly valued. While the author’s initial belief in AI was exaggerated (200% more valuable), his firsthand experiences in Silicon Valley have shown the real potential of AI, such as its ability to write novels or solve complex problems.