Halfway through 2024, where is AI going?

The big model industry is shifting its focus to actual business value, rather than simply pursuing technological leadership and parameter size. Read the article to find out what's going on.

AI entrepreneur Chen Ran discovered some "strange phenomena" in the industry.

Many customers told him that they were confused. On the one hand, the big models were updated too quickly and they couldn't figure out which one was the best to use. On the other hand, they didn't know how to integrate the big models with their business. In addition, they had no idea whether their data sets could be fine-tuned into a good big model.

The final result is that people are willing to invest in large models, but don’t know where to start. Even if they have made up their minds, they are reluctant to come up with much budget.

This in turn led to large model companies starting to raise prices and engage in a price war. "In the end, they are burning money, the price is still low, and no one uses it."

Chen Ran is the founder and CEO of OpenCSG, an artificial intelligence community and ecosystem company. He believes that the ineffective internal circulation in the large model industry is a waste. Last year, companies that rushed to make large models eventually failed to avoid falling into the trap that ofo had stepped into.

This forced industry leader and founder of Zero One Everything Lee Kai-Fu to say, "If the Chinese market is so competitive, and everyone would rather lose everything than let you win, then we will go to foreign markets."

Li Youfeng, who has been engaged in technology development for many years and started AI entrepreneurship very early, is also puzzled by today's phenomenon. "In the past, when we discussed a project, we would focus on its value, but when it comes to large AI models, people rarely discuss value, but rather discuss leading."

The overwhelming lists, various rankings, and self-promoting marketing have made this industry seem impetuous and noisy. In the end, manufacturers spent a lot of money, but their products could not be put into production, and their technology was not actually very advanced.

In late May, it was revealed that the Tsinghua-affiliated large-scale model company Xianyuan Technology had changed its leader, and its founder Zhou Bowen would leave the company. After the news came out, some entrepreneurs said: Focusing on algorithms may be a detour.

Large-model entrepreneurship is a narrow road in China, and it may be a road of no return for some teams.

Now, 2024 is halfway over, and more than a year has passed since the "100 Model Wars". Where has the industry developed to? Where will it go next?

1. Test technology: Too many people get high scores, too few can do the work

The domestic large-scale model industry is obviously "quieter" this year than last year. Last year, everyone participated and hundreds of large-scale models were released. This year, except for a few large technology companies and leading startups, most of the others have stopped.

Because everyone has discovered that no matter how loud you shout, it’s all in vain if it doesn’t land on the ground.

There are not too few large models in China, but too many, especially those that are boastful.

"Manufacturers always promote what big models can do, but they don't say what they can't do. Customers are a little misled and think that big models can do everything. They want to overturn and redo the original business, which is unrealistic." Chen Ran told "Dingjiao".

Looking back at the development of the large model industry in the past year, we will find that the first war was not a price war, or even a technology war, but a marketing war.

Marketing is to grab the attention. Holding press conferences, brushing charts, placing advertisements, and even colluded with competitors can attract more attention and make people "feel" that they are ahead. As for whether it is easy to use and its true technical strength, we can make up for it later.

Li Youfeng told Dingjiao that all the so-called self-developed large models in China are basically modified based on open source architecture, and there is no real originality or self-development. This means that the technical gap between large model companies is not large.

This is why a startup can launch a new large model from scratch in two or three months. The best example is that last year, Kai-Fu Lee's Zero One Everything released the "Yi" series of models, which were accused of using the LLaMA architecture and only renamed two tensors.

Chen Ran believes that the domestic large-scale model has not yet formed a complete entrepreneurial ecosystem. Everyone rushes in and releases a few models, which proves nothing. He uses the early days of the smart car industry as an analogy: everyone wants to build cars, tires, engines, and even windshield wipers, but the most basic batteries, electronic controls, and even wheels and seats are not ready.

From a purely technical perspective, to date, no domestic team is in an absolute leading position.

AI big models have three major elements: algorithms, data, and computing power. Domestic manufacturers have been working on algorithms in the past. When they release models, they are essentially releasing a set of algorithms and systems. Everyone competes to see whose algorithms are more advanced, whose model parameters are larger, and whose reasoning efficiency is higher. But now more and more practitioners have discovered that there are actually no barriers to algorithms. Chen Ran even bluntly said that "big models are not valuable."

"I think large enterprise-level models are meaningless. Open source enterprise-level models are enough because the most important thing is data," he said.

Data is a more scarce resource than algorithms. Algorithms can be iterated by modifying open source models and human-wave tactics, and computing power can be obtained by spending money to buy cards, but high-quality data has no channels for sale and cannot necessarily be bought with money.

Training a model is similar to training a student. Data is equivalent to teaching materials or educational resources, and the process is called pre-training. Children in remote mountainous areas and children in first-tier cities have different educational resources and training processes since childhood, and the probability of being admitted to a key university in the college entrance examination is bound to be different. In a sense, having high-quality data is half the battle for pre-training.

In the past year, the industry's standard for evaluating the quality of a large model has been through evaluation, which is equivalent to an exam. Since it is an exam, there is room for cheating, or you can get high scores by doing more questions. As a result, many large models are actually the product of "exam-oriented education" - large parameters, high scores, strong performance, but little practical ability.

Li Youfeng believes that algorithms have great limitations and are meaningless if they are separated from specific application scenarios. "For example, a model may have large parameters and strong computing power, and may perform well in math problems, but this does not mean it can generate value in actual business."

Since this year, the trend of large-scale model comparison has improved, and various "wild rankings" have also been restrained, indicating that the public is not easy to fool. The question is, if not parameters, what else can people compare?

2. Volume price: C-end dare not accept, B-end cannot afford

The most direct way for a model or a project to prove its value is to make money from the market. This year, more and more AI entrepreneurs and investors have begun to talk about business models.

There are two major categories of commercialization in the big model industry: To C and To B, which means charging individual users and charging enterprises (including governments and developers). Last year, the industry reached a consensus that it is difficult to charge To C, so we should start with the B side.

B-side enterprises are the largest customers of big models. An employee of a system integration company once told Dingjiao that they had access to Baidu's Qianfan big model platform very early and were very willing to embrace big models. However, they did not use the model because of its good effect, but simply because they were afraid of being left behind by AI. Once the model is charged, they will have to think again.

This represents the mentality of many companies: if you can get something for free, you should get it for free, but if you pay, you must see results. In Chen Ran's words, "If you want customers to spend money, you must let them see exponentially increased efficiency. Don't throw the eagle until you see the rabbit."

Li Youfeng believes that companies that really use big models focus on business data rather than algorithm indicators. "For example, if the conversion rate, click-through rate and other key indicators are improved by several percentage points, if these cannot be achieved, even if the algorithm has 1 trillion parameters and the price is as low as 0.1 cent, customers will not buy it."

In this wave of price wars in May, the price of calling large model APIs dropped by more than 90%. Taking ByteDance, Alibaba, and Baidu as examples, the price per million token inference input dropped to 0.8, 0.5, or free.

However, this was interpreted more by the market as a marketing behavior, with a bit of a clearance sale.

Jia Yangqing, founder of Lepton AI and former vice president of Alibaba, said: "It's not that APIs are expensive that no one uses them. It's because companies must first figure out how to use them to generate business value. Otherwise, it's a waste no matter how cheap they are."

Li Zhifei, founder of Mobvoi, said bluntly: "Lowering the price of API to near zero shows that OpenAI's two business models of charging consumers and charging enterprises are not sustainable in China's competitive environment."

Baidu is focusing on both the C-end and B-end. Among them, Wenxin Yiyan 4.0 for the C-end is a paid version, with a monthly subscription of 49.9 yuan. Baidu has not announced the payment rate data of Wenxin Yiyan. According to the data of the AI ​​product list, the up-and-coming Kimi, the dark side of the moon, had more visits to the web version than Wenxin Yiyan in April. Kimi did not choose to "discard his own martial arts" by collecting membership fees, but instead launched a very alternative reward function, which is not mandatory and can obtain priority rights when computing power is insufficient during peak hours.

This is still the top player in the C-end large model market, which shows how difficult it is to charge. In the final analysis, the current large model products are not good enough to use, and they are not so urgently needed. People need to be given a reason to pay.

AI startup Yuhe received investment from Qiji Chuangtan. Its initial products were all oriented to the C-end to solve various practical needs. Founder Chi Guangyao told Dingjiao that one of its main products, CopyAsk, can be used for free or for a fee to unlock more features, but more than 99% of users are using the free quota for free, and the money collected is difficult to support the company.

At the beginning of this year, Yuhe transformed itself and started to develop Agent products for the B-side. It has already won two orders. "Customers are willing to pay, and it's great to make money now." However, it will take some time to explore to completely run a certain vertical B-side business.

A very small number of companies have grasped the market demand, opened up business scenarios, and been the first to make money.

Chen Ran wants to build an ecosystem. His OpenCSG is an online community and sells software CSGHub and Starship offline. Its customers are mainly B-side enterprises and D-side developers. It has explored two monetization models: application commission and user subscription. In the future, it can also add a computing power commission model. He told "Dingjiao" that the company expects revenue of tens of millions and profit of several million this year.

3. Volume Application: No blockbuster applications have appeared, and it is difficult to implement products and projects

As big models develop to this day, practitioners are trying to make money while waiting for the emergence of blockbuster applications. Previously, we have witnessed the popularity of applications such as Miaoya Camera, Kimi, and Suno, but these applications are not considered blockbusters. Only when blockbuster applications emerge can it be proved that AI is not just a theory.

When the big model manufacturers started the API price war, some people disagreed, some sneered, and some were extremely excited.

As an independent application developer, Chi Guangyao believes that the API price reduction is a huge benefit. Before the API price reduction, he had to spend about 200 yuan per month on model debugging. Now he uses the DeepSeek-V2 model after the price reduction, and it only costs 1.11 yuan in more than half a month.

He developed several applications last year, but due to the high inference costs caused by high-frequency calls and the unwillingness of users to pay, the products have not been launched yet. Now he can't wait to make these applications "run", "If I didn't have B-side orders that I couldn't push off, I would directly develop C-side products." At the same time, those B-side orders that could not be processed before because they could not afford the API call fees can now be processed.

He believes that in the next few months, there will be a large number of explorations on the application scenarios of big models, which is likely to lead to a significant increase in application scenarios. Some high-frequency, low-logic-demand, and delay-insensitive application scenarios that used to rely on manual or engineering methods may be replaced by free big model APIs.

The implementation of big models must start with breakthroughs in scenarios. Regardless of whether the price of big model APIs is reduced, finding scenarios will become a consensus in the second half of the year.

Li Youfeng believes that more and more applications will go viral in the second half of the year, and they will find the right scenarios and cover the cost of scale based on scale output. "Everyone should work hard to find value, not price."

Among the current AI applications, two categories have demonstrated their value and received good feedback.

One is to improve efficiency. Kimi helps workers in the workplace search for information and organize documents. Zaowu Cloud uses AI to design products and marketing materials for brands. Founder Qiu Yiwu told "Dingjiao" that they used AI to make design plans for 2,000 cups for a coffee brand. Excluding model investment, the computing power cost was only 10 yuan.

The other category is entertainment, such as AI song-writing software like Suno, and AI companionship and role-playing that many startups are working on.

It is generally believed in the industry that Agent AI will be the focus of the industry in the second half of the year, and practitioners at home and abroad are all moving towards this direction.

As he explored the industry more, Li Youfeng discovered that the real difficulty in AI lies in product and engineering (referring to a series of technologies and practices for building, developing, and deploying large machine learning models). "Continuously moving large models toward product, engineering, business, and industry is the only way for Chinese startups to grow."

Once the engineering problem is solved, the model is not important. When using the product, users do not care which model is used at the bottom layer or what proportion of the product is self-developed. The effect is better than anything else.

The current situation is that the big model cannot guarantee the effect 100% when it is applied. Take the big language model as an example, the problem of "nonsense" has not been overcome. "This uncontrollable state makes it difficult for it to play a big role in production, and it still needs time to polish. For most B-side customers, how to keep up with the development and evolution of the community and big model ecology is what they should pay attention to." Chen Ran said.

He believes that the industry was too optimistic about big models before, and there was always an unrealistic fantasy of "training a big model to change the world". In fact, big models have not yet truly formed productivity. There is still a transition process between AI 1.0 and AI 2.0. "It is about how to empower the existing system through AI first, rather than completely denying it."

In the process of climbing up, some companies will inevitably fall and be eliminated, especially those that do not yet have the ability to generate revenue. In a report, Stanford University's Institute for Human-Centered Artificial Intelligence said that global investment in artificial intelligence fell for the second consecutive year in 2023.

In China, technology giants represented by Alibaba and Baidu are still continuing to invest. For example, Alibaba spent US$800 million to invest in Kimi's parent company, Dark Side of the Moon, holding a 36% stake. Market rumors have it that Tencent is also in talks to follow suit.

However, it is increasingly difficult for us to assess how much these big companies have done and what roles they have played in promoting original innovation and product implementation. On the one hand, big companies build big models themselves, and on the other hand, they invest in almost all the star startups on the market, using equity to bind their competitors. Even this wave of API price cuts praised by developers was not initiated by big companies, but by a private equity giant called Huanfang Quantitative and the startup Zhipu AI, followed by big companies, which is full of passivity and marketing flavor.

In the second half of the year, the industry will still be very competitive. After all, the domestic big model has been driven by competition from the beginning. Perhaps we can create a few winners, and the hit application is not far away.

Author: Liming

WeChat public account: Dingjiao (ID: dingjiaoone)