news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Text | Small is Big Entrepreneurship Partner

Since January this year, we have held a large model application salon in Shanghai every month. Here are some of my observations:

About the development in the past year:

The cognition is aligned, flowers bloom everywhere, a small profit is made, and we look forward to the hot products.

At the beginning of the year, some speeches at the conference would say things like "our company has five years of experience in artificial intelligence," intentionally or unintentionally mixing LLM with previous AI. Now, everyone can be very frank about which scenarios are using previous AI technologies and which are using models. They can also clearly state which are based on the models and which scenarios have been fine-tuned or retrained. This is a good phenomenon, as it allows everyone to more directly enter the areas that should be discussed.

At the same time, we are seeing more and more scenarios where LLM is actually used, including the combination of content generation and specific work scenarios. New RPA products that use LLM technology are actually being implemented in areas such as human resources interviews, screening, and job training.

For projects that have already been implemented, some parties A already have relatively quantitative data to evaluate the output of LLM, and some parties B have begun to say that they have signed contracts or received project funds. Most of the projects undertaken by parties B seem to be in the hundreds of thousands, with occasional projects worth a few million. Income from this direction has not yet become the main support for the company's income.

At the beginning of the year, people were asking, and this question is still being asked now, that is, what are the hot-selling software for large-scale 2B models? It seems that there is no such software yet. Maybe it will be a long time before there is one? What have been the hot-selling 2B softwares in the past years? Office automation? Financial ERP? CRM? DingTalk Feishu? Let us continue to look forward to it.

About product form:

Chat interface, traditional interface, traditional + chat = Copilot, and more?

People have different expectations for big model 2B software, some are pessimistic, while others seem to be overly optimistic. We visualize this expectation and think about how everyone in the enterprise will feel about this new big model-driven enterprise software.

The first time the big model came into people's view was as a chat interface. Gpt3.0 had demonstrated capabilities that were very similar to the current ones, but because there was no chat, only API, it did not attract more attention and did not break through the circle. Although the chat interface and search look very similar, you will feel the essential difference once you have used it. We have reason to believe that the new generation of enterprise information system infrastructure with chat as the interface and available to everyone in the enterprise will be popularized in all enterprises like office software, email, enterprise WeChat/DingTalk/Feishu. Colingo.ai, incubated by the Small is Big entrepreneurial partner, focuses on this market.

In traditional enterprise software, some backends have already used various AI algorithms, and some of them have been replaced by large models to better solve problems, or large models have been used to handle problems that were previously impossible to handle with algorithms. These things have already happened quietly, but the end users' perception is relatively weak. There will be more and more such scenarios, but it may not be easy to produce popular software.

The combination of chat interface and traditional GUI software is Copilot. As an early close partner of OpenAI, it launched Copilot for Office, allowing users to better use traditional software through the chat interface. It seems that this will also be one of the main interface forms for a while?

What new interface forms are there? How do digital employees collaborate with other employees? There is still a lot of room for discussion in this regard.

About the technical route of 2B software service providers:

• Top three: Arrangement, rag, agent
• The bottom three: computing power, infrastructure, training and fine-tuning

Knowing too much is actually a burden. This phenomenon does exist in the era of big models. Not to mention that the concept of "big model thinking" may be difficult to define, but from the process of various software service providers starting with big model application, it is still reflected.

The difference between the upper three routes and the lower three routes is not just a few GPUs. The talents, workflows, delivery cycles, iteration speeds, and overall comprehensive costs involved are not on the same order of magnitude.

The first reaction of the technical departments of many companies is to train the model, buy cards, and recruit people. Then they may fall into a quagmire, and feedback to the CEO becomes that the big model is a very heavy task and requires a large investment. There is no progress within the company, and externally, customers keep asking what they can do for them with the big model, which puts them in a dilemma.

The upper three paths can solve most of the problems involving large-model applications. Application companies should enter the lower three paths at least after completing the upper three paths and making sure that the lower three paths can really and significantly improve the results.

Not allowing the word "training" to be used in large-model application development may be an overreaction, but at this stage, for most application development companies, aiming at the top three routes may be the fastest way to achieve results.

There are two more specific issues that I have discussed with many companies and would like to share with you:

Regarding what customers choose to start large model applications:

"Advanced enterprises use Feishu", we need to find those clients who already have "big model thinking". As a service provider, it is difficult to brainwash the clients. We need to find clients who are happy with each improvement in results. Before they start, there are 10,000 problems ahead. We should avoid such users.

About what model to choose:

Start with the best model you can get. If you can’t do it, it means that this generation of model is not good enough. Just wait.

Once it is done, use the most realistic model to optimize it. Tell the client the difference between the optimization levels. If they can accept it, go ahead. If not, wait.

Don’t consider the cost of the model. Today’s price is completely incomparable to that of six months ago, and the capabilities six months from now will definitely be higher than today’s. To develop large-scale model applications, you must have a basic belief in the “Moore’s Law of Large Models”. Instead of being obsessed with optimizing the current model, it is better to embrace time as your friend.

To summarize the judgment of opportunities:

In the market of large-model 2B applications, there are new scenarios, new forms, and new companies, and old scenarios, new solutions, and new and old competition.

There is no doubt that it is an opportunity for new enterprises, and the product form is far from mature. The iPhone was released in mid-2007, and what mobile applications did you remember at the end of 2008? The big model at the end of 2022 is only a year and a half away. Of course, enterprises need to survive, but survival cannot be the goal. There are still many opportunities to explore. If it's not you, then it's others.

The opportunity also belongs to software service providers in various industries. 2B services ultimately need to solve user problems, and you are the ones who understand user needs best. In the application field, competition in the era of big models has greatly weakened technical barriers. The ability to transform user needs into problems suitable for big model processing has become the most important competitive point, and this is undoubtedly the advantage of industry software service companies.

A reshuffle is bound to happen, no matter how long you have been in this industry. Whether you are the first or the last in the industry, you may be eliminated in this competition, leaving a few places for new entrants of large-model native companies, and the rest will also be re-ranked in this technological revolution.

The decisive battle will be this fall, this day of the year.

The first re-ranking will happen in the fall of this year. Most companies will not have a special budget for big models in 2023, and individual purchases are diverted from the budget of other IT projects. There will be some budget for POC related to big models in 2024, and a lot of IT budget in 2025 will definitely be related to big models. Whether you can make your solutions and your company's vision in the fall of this year will determine how many projects you can get in 2025 and your position in the first qualifying battle.