Practicing the large model in a muffled voice, Byte launched a surprise attack

21st Century Business Review

2024-05-17 06:36Published on the official account of Guangdong 21st Century Business Review

Title picture source: Internet

Written by丨He Jipai Editor丨Yan Ziwei

On the battlefield of the large model, Byte carried the "bean bag" to catch up with the evening set.

On May 15, ByteDance's cloud service platform, Volcano Engine, released a large model of bean bags. In its debut, it released 9 models in one go, and offered lethal weapons at a low price.

According to Tan Cheng, president of Volcano Engine, the pricing of the main model of bean bags in the enterprise market is only 0.0008 yuan/1000 Tokens, and 0.8 centimeters can process more than 1,500 Chinese characters, which is 99.3% cheaper than the industry.

As soon as the price was announced, the venue was in an uproar, and there was a "wow" sound.

In a media group interview after the meeting, Tan Cheng said that the ultra-low pricing comes from confidence, "We will not do this with losses." ”

Over the past year or so, Byte has disclosed very little about AI. Behind the low profile, its internal catch-up has accelerated, and multiple departments have gone down the field to develop AI products.

As soon as it came out of the mountain, Byte set off a high-profile price war for large models.

Price "Butcher"

Up to 99% down, the byte knife is cut directly to the ankle.

"The model has to work well and be cheap enough." In the price announcement link, Tan Zhi made it clear as soon as he came up, and directly ordered Baidu and Alibaba.

He mentioned that the pricing of large models is generally based on thousands of tokens as the main billing unit.

Among the main models below 32K, OpenAI's GPT-4 is priced at 0.42 yuan/1000 Tokens, which is cheaper for domestic manufacturers, and Baidu Wenxin Yiyan and Ali Tongyi Qianwen are about 0.12 yuan/1000 Tokens.

The main model of bean bags is priced at 0.0008 yuan/1000 Tokens, which is 99.3% cheaper than the industry.

To convert, that is, users can buy 2,400 Tokens from GPT for 1 yuan, and the bean bag with the volcano engine is 1.25 million Tokens.

Tan Bei made an analogy, which is equivalent to the amount of text for processing 3 copies of "Romance of the Three Kingdoms", which only costs 1 yuan.

After the meeting, a large number of questions were asked about the price. Tan Cheng emphasized that the pricing lies in two points, one is what needs to be done, and the other is what the volcano engine can do.

At the same time, there are many technical means to help reduce costs, such as the use of distributed inference, large-scale hybrid scheduling, etc.

On the question of "whether to exchange losses for income", Tan Bei responded clearly, "It is not sustainable to exchange losses for income, and we will not take this path." ”

Bytes take the lead, and the price of the large model is absolutely good from the perspective of the enterprise.

An entrepreneur told 21CBR that his company does not develop its own large models, and is open to the call of various large model platforms. "The big factories fight to the end, and only a few will definitely survive."

Team Battles

"Don't take bean bags and don't use dry food", Byte's press conference this time did not take the usual path.

The first is to fight a price war when it comes up. It is reported that the company attaches great importance to this pricing announcement, and for the sake of confidentiality, the staff who participated in the rehearsal do not know the price.

The second is the "three-board axe" commonly used by large model manufacturers, and the number of enterprise customers, parameter scale, and list score are not mentioned.

"We're always thinking about what makes a good model. Is it the first place in various evaluation lists, or the largest number of parameters, or is it claiming to surpass GPT-4 on PPT? It may be that these are one of the factors, but not enough. ”

Tan to be treated

Tan said that the large model can only be implemented in the real scene, and the more people use it and the larger the number of calls, the better it will be.

It disclosed that the Doubao model completed its self-development a year ago and has been connected to more than 50 businesses within Bytes, including Douyin, Feishu, etc., processing an average of 120 billion Tokens text per day and generating 30 million pictures.

Bytes, which has always mass-produced APP, adopted the team battle method in this release, and a large model family bucket was served for different scenarios to choose according to demand.

The 9 models unveiled, including the general-purpose model Pro, the general-purpose model LITE, the speech recognition model, the speech synthesis model, and the Wensheng diagram model.

Lite, a general model of bean bags, is more economical and suitable for some scenarios that do not require high model capabilities, such as the interaction of automotive intelligent cockpits.

Compared to Pro, the cost of 1,000 tokens in the Lite version is reduced by 84% and the latency is reduced by 50%.

Compared with the above two general models, the model featuring voice is a highlight of this conference, including speech recognition, speech synthesis, voice reproduction and other directions.

Coincidentally, just 1 day ago, OpenAI launched the GPT-4o large model, and the new feature that can read user emotions has sparked heated discussions.

In Tan Bei's view, let the large model express real feelings, and the dialogue between people and AI will be like a real person, and the experience will be better.

"Whether voice is done well or not has a great impact on the interactive experience. A model is smart, but it speaks dryly, like a robot. We put a lot of effort into speech. He said.

Innovation anxiety

In other large factories, the number one leader personally commands the "supervisor" model, and Byte is low-key to the point of almost mystery.

Byte has a large amount of investable funds and is extremely cautious.

In April last year, Tan Bei mentioned the development idea of large models, saying, "Volcano Engine does not make large models, but first serves domestic companies that do large model entrepreneurship." ”

At that time, a number of large manufacturers and entrepreneurial start-ups flocked to the large-scale model track, but Volcano Engine showed the role of "AI shovel seller", focusing on providing computing power and service support.

The top is not without anxiety.

At the beginning of the year, ByteDance CEO Liang Rubo mentioned that it was not until 2023 that the company began to discuss GPT. The large-scale model startups that do well in the industry were all founded from 2018 to 2021.

In the past year, Byte chose to hibernate and increase research and development.

In June last year, news came out that Byte was testing an AI conversation product, codenamed Grace.

Two months later, its AI dialogue product "Doubao" App opened public beta.

Bean bag is a large-scale model application that Byte focuses on.

In the Apple App Store and major Android application markets, the number of downloads of the Doubao APP ranks first among AIGC applications in China.

According to Zhu Jun, vice president of product and strategy at ByteDance, more than 8 million agents have been created on Doubao, with 26 million monthly active users.

In addition to the AI dialogue assistant "Doubao", Byte has also created the AI application development platform "Button", the interactive entertainment application "Cat Box", and the AI creation tools Star Painting, Instant Dream, etc.

"A year ago, the domestic large-scale model technology was average, and I think it was not time to talk about application breakthroughs. Now, in fact, it has reached this stage. Tan Cheng said.

Entering the large-scale model track at a low price may be a signal for this company to start the charge.

View original image 368K

• 

• 

• 

• 

• 

• 

• 

• 

•