Several Certain Trends in AI Development Over the Next 6-12 Months

The recent views on AI in the capital market are quite interesting. From cooling down in the second half of last year, to questioning its output ratio at the beginning of this year, and then to regaining confidence recently, the past few months have been like a rollercoaster ride. Although concerns haven't been completely dispelled, the worst should be over. Regardless of how capital views it, the development of AI is still predictable. If we learn from history, we can easily predict what will happen in the next year or even the next six months. As someone working in the technology industry, I'd like to share some of my thoughts.

Models are stabilizing and AI applications are booming.

In the past two years, many startups have targeted AI applications, but after working hard for a long time, they found that their hard-earned products could be wiped out by a major version upgrade of the large model. AI applications are based on the optimization and in-depth exploration of the capabilities of large models. If the large model is upgraded too frequently and the improvement in capabilities is too large each time, companies will not be able to develop applications properly.

The large model progressed from the capabilities of an elementary school student to a middle school student and then to a high school student in less than three years. It's foreseeable that it will soon become a "college student," and this college student will likely become a "platform," after which development will plateau – making this a good time for AI applications. This is similar to the development of the internet and mobile internet. During the internet revolution of the 1990s, everyone was eager to showcase their abilities, but the real explosion and monetization occurred in 1996 and 1997. It only truly got on track after the dot-com bubble of 2000, and many of today's well-known companies took off during this period. The mobile internet, which is closer to us, follows a similar trajectory: the release of the iPhone in 2007, the gradual maturation of Android, and the explosion of various apps in 2010 – a development period of about three years. Many of today's well-known companies started around 2010. This pattern isn't just due to the technology development cycle, but also reflects the market's acceptance of new things, the entry of startup teams, and the shift in investor behavior.

The explosive popularity of OpenClaw at the beginning of the year already demonstrated the market's desire for AI applications, and recently it's become very clear that more and more AI applications are starting to be implemented. I think it won't be long before more and more vertical applications emerge, and a large wave of startups will rise up.

The Embedding of Enterprise Workflows Has Exploded

Currently, most AI applications are based on chat windows, with a question-and-answer format. This model isn't particularly efficient, but it addresses complex user needs. However, as AI develops, the demands on it will become more specific—shifting from solving broad, general problems to addressing more niche areas. This is similar to the invention of the steam engine, which could generate vast amounts of general power, but how to utilize that power to solve real-world problems requires exploration by professionals in various fields.

It's a consensus that businesses can save costs by leveraging AI, but how? Currently, AI cannot function without humans (and we don't foresee a future where AI can operate independently), making it essential to set a checkpoint at each stage of the process. If we compare a company's operations to an assembly line, the most practical upgrade would be to divide the assembly line into smaller segments, each with a central control unit. A person would operate this control unit to confirm the AI's results before proceeding to the next step. This segmentation and integration process is very time-consuming and labor-intensive, but once implemented, work efficiency will increase exponentially, while the number of employees required will decrease significantly.

These enterprise-level services will become highly sought-after by large companies. However, just like website development in the past, the market is so large that it's impossible to divide it all. Even if large companies take all the profits, the remaining amount will be enough for small businesses to thrive. This trend will continue for several years. In the foreseeable future, many jobs that require human labor will be replaced by AI, and companies will only need a few people, or even just one senior staff member, to operate the central control system.

Token usage surges, price drops

With the explosion of AI applications and their embedding into workflows, token usage will surge. In fact, we have already seen this trend. However, I think this year's surge in usage is just the beginning and is far from enough. This may be like the bandwidth and data traffic of dial-up internet back then. We could not have imagined that the internet would have such high traffic and that our homes would have such high bandwidth and speed.

Many people have no concept of tokens, assuming that using AI is completely free, just like the internet. However, the amount of tokens used by AI is often beyond imagination. Every interaction and conversation you have will use hundreds or thousands of tokens, and these tokens are not cheap. The OpenClaw crayfish that went viral last month is a case in point. Many people excitedly set up their machines and let them do some seemingly magical but actually meaningless things, only to suddenly find that the bill was tens of thousands of dollars.

Currently, AI applications are limited to "query-and-answer" functionality. However, once 2C applications explode and 2B workflows are embedded, token usage will grow exponentially. This explains why major companies are investing heavily in computing centers – this will be the biggest growth driver for the next ten to twenty years. The number of tokens is essentially computing power, and computing power is essentially electricity. Computing power is influenced by chips, while electricity is a fundamental infrastructure. In this respect, the US power grid is much weaker than China's; however, the US has many electricity generation technologies independent of the power grid, and these technologies will flourish in the future.

Unlike a market economy, the explosive growth in token usage will not lower its price (as mentioned before, the price is affected by computing power and electricity). However, the establishment of computing centers and the upgrading of power infrastructure will greatly reduce the cost of tokens, thereby lowering their price. Roughly speaking, the cost of using AI should not increase much, which will in turn promote the application market (enterprises will definitely use it because it can save them a lot of costs).

From horizontal development to vertical development

The current development of AI is horizontal, constantly expanding its general capabilities. However, once a platform is established, many vertical development opportunities will emerge. This is not difficult to understand. When a platform is stable, AI applications are used to solve a specific type of problem. As more applications solve different problems, they naturally become more specialized. This is determined by the market. Customers need solutions that address their pain points, not a jack-of-all-trades.

Currently, AI's penetration into other industries is very limited. As it is embedded in workflows, various industries will gradually have pain points that can be solved by AI, which will be captured by entrepreneurs, leading to the emergence of more AaaS (AI as a service) companies. These events can be fully referenced from the development path of the Internet.

Postscript

Tonight, a sudden inspiration struck me, and I wanted to share my insights. This world is inherently unpredictable; many things, once understood to their essence, reveal a clear path for development. However, simply knowing this isn't enough; timing, action, and luck determine success. The world's element of chance is both alluring and regrettable.