Article reprint source: AIGC

Source: Shidao

Image source: Generated by Unbounded AI

Peter Thiel, a famous Silicon Valley investor and founder of PayPal, an internationally renowned payment tool, once said: "Competition is for losers. If you want to create and capture lasting value, build a monopoly."

This phrase is the ultimate expression of "Winner Takes All," which means that if a product or service is just a little bit better (e.g., 1%) than your competitors, you will get a disproportionately large share of the revenue (e.g., 90-100%) in that category, leaving your competitors far behind.

This phenomenon is reflected in many industries, especially the technology industry. Globally, IBM has dominated computing for decades; Microsoft dominates the personal computer market; Amazon still single-handedly dominates e-commerce. Obviously, a major feature of the Internet era is "winner takes all, and losers are always weak."

It is important to clarify this issue because it will change our investment logic: if the traditional Internet approach of "burning money for subsidies - eliminating the second place - monopolizing the market - leveraging network effects" no longer works, then investors who have experienced the "Thousands of Group Wars" era may also need a new investment methodology.

In order to get the answer, the Shidao investment research team referred to several external articles, whose authors included A16Z partner Benedict Evans, Lightspeed Venture Partners partner Guru Chahal, and others, trying to sort out some similar or contradictory views for everyone to think about.

Winner takes all: Representative Andrew Ng

Overall, the virtuous cycle model of artificial intelligence introduced by Andrew Ng provides the underlying logic for the "winner takes all" approach.

Initially, AI products are built with limited data. Then, as they interact with users, they collect more and more data every day. The foundation of machine learning is data—lots of it.

More data = more accurate models = better products = more users = more data

This virtuous cycle formula is considered an important factor in the winner-take-all market for artificial intelligence. The combination of big data and machine learning amplifies network effects and returns to scale, which once again reinforces the dominance of technology market leaders, meaning that companies that are already large and have a lot of data will become even more powerful.

In China, data barriers are also a wall facing emerging companies. High-quality Chinese corpus data is a big challenge for startups, and data accumulation requires time and experience. For companies like Baidu that have accumulated data through search and other Internet and IoT applications over the years, they are already several steps ahead from the beginning.

No winner takes all: A16Z

Data is crucial, but A16Z partner and well-known analyst Benedict EvansDoes has a different view on the role of data in actual work.

Evans Does in the article “Does AI make strong tech companies stronger?” points out that although machine learning requires a lot of data, the data you use must be very suitable for the problem you are trying to solve.

GE has a lot of telemetry data from gas turbines, Google has a lot of search data, and American Express has a lot of credit card fraud data. But you can’t train a model on turbine data to spot fraudulent transactions, nor can you train a model on web search data to spot gas turbines that are about to fail.

Every model you train should do only one thing.

This is very similar to previous waves of automation: just as a washing machine can only wash clothes but not dishes or cook, and a chess program cannot pay taxes, a machine learning translation system cannot recognize a cat.

The application you build and the dataset you need are strongly related to the task you are trying to solve. (Although this is a moving target, there is research trying to discover how to make machine learning models more easily transferable between different datasets).

This means that Google is getting better at being Google, but it doesn't mean that it's getting better at other things.

Some industries will, some won’t: it depends on the specific vertical field

So, in vertical fields, can leading companies seize the entire market by relying on their far-leading data advantages?

EvansDoes think the situation will become more complicated.

Who owns the data, how unique is the data, at what level is the data unique, and where is the right place to aggregate and analyze the data. The answers to these questions will be different for different business units, different industries, and different use cases.

Let’s assume that if you are building a company to solve real-world problems with machine learning, you will face two basic data problems:

1. How do you get the first batch of data to train your model to get your first customer?

2. How much data do you need?

The second question can be broken down into many questions:

Are you solving a problem with less data that is easily available? (But available to many competitors).

Or do you need more, harder-to-get data to solve the problem?

If so, is there a network effect that could benefit from this, where one winner gets all the data?

Does the product get better indefinitely with more data, or is there an S curve?

It all depends on the situation.

Some data is unique to a company or product, or has strong proprietary advantages, such as General Motors' turbine telemetry technology. But this may not be very useful for analyzing Rolls-Royce turbines.

Some data can be used for use cases in many companies or even many industries. Many startups are born to solve common problems of many companies or different industries, and the data here has a network effect.

But there are also cases where after a certain point in time, the vendor doesn’t even need more data because the product just works.

EvansDoes believes that this situation has already happened in many startups. For example, Everlaw, a company invested by A16Z, has developed a legal software that can perform sentiment analysis on one million emails without the need to train it with the client's specific litigation data as raw material.

In an even more extreme case, a large vehicle manufacturer is developing a more accurate tire blowout detector through model training. This is a model trained based on a large amount of tire data. But obviously, this data is not difficult to obtain.

That is to say, the popularity of machine learning does not mean that Google has become more powerful, but that a variety of startups can use this cutting-edge technology to build an application and solve a problem faster than before.

There won’t be more “AI” startups in the future, they’ll be industrial process analytics companies, legal platform companies, or sales optimization companies.

EvansDoes compares machine learning to SQL (Structured Query Language).

In the past, if you didn’t use SQL, you would fall behind. For example, one of the key factors for Walmart’s success was using SQL to manage inventory and logistics more efficiently.

But today, when you start a retail company and say “…we’re going to use SQL,” it doesn’t make the company look more valuable because SQL is already part of everything and then it gets lost in the discourse.

The future of machine learning will be no different. ‍

How to invest in the era of "big models"? At least the Internet logic does not work anymore

The Shidao investment research team believes that regardless of whether there will be a "winner takes all" situation, at least the investment logic of the Internet era will no longer work in the era of artificial intelligence.

The core logic is that in the Internet era, "traffic" is free, which is why there is the concept of "network effect": that is, with the total operating cost unchanged, the more users there are, the greater the value of the network. This is what is called "all industries are suitable for redoing with Internet thinking."

However, the difference in the era of big models is that computing power has a cost. Therefore, every additional user requires actual computing power, and there is no network effect. This makes subsidies meaningless. The more new users you have, the less money you make.

In addition, the current large models have problems such as high cost of use, large inference delay, data leakage, and inaccuracy in professional tasks. In comparison, the advantages of some smaller and specialized (adjusted + refined) long-tail models have also emerged.

Therefore, even though most technologies can play a role in wealth accumulation and artificial intelligence giants can indeed accumulate a lot of wealth, the total amount of wealth will become limited due to computing power costs and the inability to occupy the entire market.