AI startups in the direction of big models and intelligent agents - three key suggestions for your reference

The day before yesterday, I happened to visit the head of a venture capital institution. We were originally discussing some project cooperation matters, so naturally we talked about the recently popular AI entrepreneurship.

The person in charge also said that after the Chinese New Year, he basically meets 2-3 AI startup teams every week to listen to their products, solutions or business plans. So after talking with him for 2-3 hours, I felt that this person in charge actually has a deep insight into the AI ​​startup market, including his own deep understanding.

Why do I say so?

Because although many venture capital institutions are also eager to find good AI startup projects, the money of venture capital institutions is not picked up for free, so they are relatively cautious when choosing AI startup teams. To put it bluntly, even if AI and DeepSeek are popular now, it is not easy for startup teams to win investment from venture capital institutions in the field of artificial intelligence.

Today's article will summarize some key thoughts that I discussed with venture capital institutions about the direction of AI entrepreneurship. The focus is on the following three points for your reference.

1. Volume bottom layer model or volume top layer scenario?

The first point is that for AI startup teams, you should not focus too much on the underlying technology and the underlying big models, but should focus more on the upper-level AI intelligent agent applications.

Some startup teams may feel that my technical foundation is very good, but if you want to make the underlying large model, including some optimization of the large model algorithm, it will actually cost you a lot of money and a long time period.

After DeepSeek became popular after the Chinese New Year, people may have the illusion that it is not too difficult to optimize the underlying large model or algorithm. But in fact, this is still quite difficult. Even if you stand on the shoulders of giants, you still need a lot of time accumulation, cost and personnel investment. Including everyone seeing DeepSeek is popular, in fact, the Magic Square team itself has accumulated algorithms and technologies in this direction for quite a long time.

Moreover, if you focus on making underlying models or algorithm optimization, you will face competition from giants. For example, if you are working on general models, others are already working on deep reasoning models. You have made R1 version, and DeepSeek may have released the latest R2 version. You are always catching up, but it is actually difficult to surpass. And you will find that a lot of your early technical research results are often worthless later.

So you should not focus on the underlying model, but should focus on upper-level industry-oriented AI intelligent application. However, this requires you to have a deep insight into a certain business segment of a certain industry.

In other words, it is not that easy to develop AI agents or AI applications. The difficulty is not in the technology of agent development. The real difficulty is that you need business experts who understand a certain industry and a specific business field to work with you on this AI application. Only in this way can you highly condense the business experts' original historical experience into your agent application. This is the first key point I want to talk about.

2. Data training - public data or private data

The second key point is that when AI startups choose related products or directions, they must choose directions where they can easily obtain data on the Internet to train and tune their models, rather than looking for private data in too many niche industries where it is difficult to obtain data.

Similar to the medical industry, government affairs industry, transportation big data and other industries. In fact, these industries themselves have a wide demand for the application of artificial intelligence, but this is often not a good direction for AI startup teams.

Because it is quite difficult to obtain these private data in these directions. If you cannot obtain private data, it will be difficult for you to conduct in-depth pre-training or tuning of the model. On the contrary, it is similar to what we often say, similar to the financial statement analysis of listed companies themselves, similar to the generation and production of some intelligent courseware in the field of smart education, which are actually some good directions.

Because it is relatively easy to obtain these public data, after we have chosen the direction, we must be more aware that sometimes we are talking about big models, and sometimes we are talking about upper-level intelligent agents. The big model is often the underlying skeleton, and the intelligent agent itself is the upper-level blood and flesh. What I need is not a simple underlying skeleton, but something with flesh and blood . If you develop an intelligent agent with a simple shell, what is the difference between it and a skinny person standing in front of you?

This is also the second key point I talked about, that is, to make a good AI agent or AI application for vertical industries, in addition to business experts, you also need to be able to easily obtain data from the Internet, train your application through the data, and integrate the trained rule experience into the agent as the core knowledge base of the agent plug-in.

3. Direction selection - is it better to go to the blue ocean or the red ocean?

The third key point is that when we start AI businesses, we always want to find new directions and new blue oceans, but in many cases, the red oceans with fierce competition may be the direction you can focus on. Because if this segmented direction is a red ocean, at least it means that others have helped you explore the path in this segmented direction, and it has huge market space and market demand.

So in these red ocean directions, you can still do one thing to the extreme, similar to an example I used to talk about often.

For example, the natural language query ChatBI direction. I also talked about this at the Huawei Cloud conference I attended at Huawei Songshan Lake in June last year. Huawei also has some similar natural language to SQL products, but the accuracy may only be 90%.

In the manufacturing field, we do quality analysis and defect prediction of production data. Now we also use a large number of AI models, but the accuracy may only be 85%.

So if you can achieve an accuracy of 95% or 99% in these segmented industries and directions, and make it sufficient for commercial use, then this product will still have considerable significance and marketing value.

Therefore, whether to choose the blue ocean or the red ocean must be analyzed based on the actual situation. If you think you have good enough technology, it is easier to achieve results and promote it quickly by choosing the red ocean market that has been verified by the market.

Okay, that’s all for today’s three suggestions on AI entrepreneurship.