In November 2022, OpenAI launched ChatGPT based on GPT-3.5, which ignited a wave of large models and caused a global AI race. Many leading Internet companies and artificial intelligence companies at home and abroad have basically released their own large models, and there is a "100-model war" in China.
According to the statistics of the AI Algorithm Filing Center, as of March 28, 2024, the number of large models filed through algorithms in China has reached 117. Among them, there are many familiar Baidu "Wenxin Yiyan", Alibaba's "Tongyi Qianwen", SenseTime Technology's "Daily New", and the ByteDance "Doubao Model" (original name: Skylark) that we are going to talk about today.
Reveal the family structure of the bean bag model
After nearly a year of iteration and market verification, the bean bag model has finally officially opened its external service today. At the 2024 Spring Volcano Engine Force Motive Power Conference, ByteDance unveiled the mystery of the bean bag model, according to reports, the bean bag model includes nine models, including general model pro, general model lite, speech recognition model, speech synthesis model, and Wensheng diagram model.
Source: ByteDance
In terms of specific applications, ByteDance has created a series of applications such as the AI dialogue assistant "Doubao", the AI creation tool "Instant Dream", and the AI application development platform "Buckle", and connected the large model to business segments such as Douyin, Feishu, and Giant Engine to improve efficiency and optimize product experience.
According to Zhu Jun, vice president of product and strategy of ByteDance, the Doubao app ranks first in the number of downloads in AIGC applications in major Android app markets and Apple App Store, with 26 million monthly active users.
Previously, Xiaolei specially experienced the AI creation tool that takes picture and video generation as the core selling point, and has a deep understanding of the strength of the bean bag model. Although the dream still needs to be optimized in terms of video generation, the picture generation has reached a good level, which makes Xiao Lei full of expectations for the bean bag, what kind of level is the bean bag that ByteDance has high hopes for, you might as well follow Xiao Lei to take a look.
The bean bag is light experience, and the whole application emphasizes extensibility
Doubao has an application entrance on the mobile terminal and the PC side, and Xiaolei found that the original Doubao has a layout in the Windows/Mac desktop client and browser plug-ins during the experience. In order to facilitate the experience, Xiaolei chose the PC web terminal as the experience object.
Source: Bean Bag
Entering the homepage of the web terminal, Doubao gave priority to recommending three different types of AI applications to me, namely AI search, PDF Q&A, and image generation, which may be the most frequently used application by users. In this case,Xiao Lei found Kimi, who is also a positioning intelligent assistant,Let's take a look at the quality of this AI search and PDF Q&A,Xiao Lei has already experienced image generation on the dream,This time I won't repeat the experience,Both are connected to the bean bag model,The performance should not be much different。
Under the strong dialogue, I believe that everyone will have a more intuitive feeling of the bean bag model.
Source: Bean Bag
AI search: each has its own strengths and weaknesses, and bean bags win in efficiency and problem extensibility
Recently, "Singer 2024" has become popular all over the Internet, and the highly topical competition between Chinese and foreign singers has aroused heated discussions among many netizens.
Doubao took a few seconds to give an answer and search source, the content of the answer did not contain common sense errors, the player information and rankings were very clear, and the ridiculous Internet memes could be accurately identified and explained. Extensibility is one of the essences of AI search, and this is also demonstrated in the answer below.
Comparatively speaking, Doubao summarizes the answers we want to know in very concise sentences.
Source: Bean Bag
Kimi is relatively better in this area, explaining the questions in more detail, and marking the source of the information in each answer, which can save users a lot of time looking up the source of information. In terms of generation efficiency, Kimi's performance is not as good as that of Beanbao, and it took about 10s to give an answer, and there is no extended question of the question, which is not a small problem for an intelligent assistant.
Source: Kimi
PDF Q&A: Bean bag surprise, Kimi stable
Long text processing is Kimi's strong point, and in his previous experience, Lei asked him to summarize the PDF file of the book "Too Noisy Loneliness", which has a word count of around 100,000. This time, the processing object of the bean bag is still it, look at the same instruction to process the same PDF file, and what kind of answer can the bean bag hand in.
Source: Bean Bag
Source: Kimi
Kimi divided the main plot of the article into sections, and each paragraph provided a summary of the paragraphs, and the beginning and end of the answer gave a precise answer to the content and moral of the article; The bean bag is expressed in the form of an article, condensing the content of the article into a short story, to the effect that it is basically the same as Kimi, and finally giving his own understanding. Interestingly, Doubao still follows the good habit of extended search, and 3 article-related searches are provided below.
In general, while accurately identifying the content of the PDF file, the two summarize the content, and the key points of the article are basically mentioned. The difference lies in the processing logic of the large model, which leads to a difference in the form of the content of the answer, there is no superiority or inferiority, and the user uses this function only to get an accurate answer. However, due to the limited information, Xiaolei did not have time to prepare a PDF with a larger amount of text, so he could not test the limit of the text processing ability of the bean bag.
In addition to common AI applications, Xiaolei found that Doubao also hides many interesting agents. There are role-playing, copywriting assistants, and various tests and other agents, and the number and categories are enough to cover many scenarios such as life, work, study, and creation. Users can also choose to create their own AI agent, and freely set the avatar, name, persona, and permissions. Of course, Xiao Lei has done an in-depth evaluation of the bean bag before, and tested the strength of the bean bag in various scenarios from multiple dimensions, and interested friends can click to check it.
Source: Bean Bag
Previously, Xiaolei created an agent with a built-in AI voice on Wenxin Yiyan, and the Doubao APP already has a similar application, but the Doubao web side has not seen the relevant settings, and perhaps the follow-up will be combined with the speech synthesis model in the Doubao large model family, the voice replication model, and the speech recognition model to iterate, so that the AI agent on the web side is more anthropomorphic.
The price of large models has entered the "centith era", and the implementation of AIGC application scenarios has accelerated
In the past year, the Doubao model has been widely used in more than 50 businesses and scenarios within ByteDance, and many users have found Doubao when using Douyin and Feishu. It is understood that since the launch of the bean bag model in August last year, it has processed an average of 120 billion Tokens text per day and generated 30 million pictures, and the huge internal usage is to better realize external services.
The joint release of large model + application products is the habit of most large model players, while the bean bag large model is just the opposite, and it will be officially released after one year of use. Maybe it's to accumulate a larger amount of data usage, or to make a more perfect debut, in short, ByteDance has a clear large-scale model strategy, and it will not be adjusted at will because of external comparisons.
At this press conference, ByteDance did not release any list scores and parameter scales, but emphasized another important factor in accelerating the landing of the bean bag model: price. The price of the main model of bean bag 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, and a simple conversion, 1 yuan can process 1,250,000 tokens, which is much lower than the processing cost of other large models such as GPT4, ERINE4.0, and Qwen 2.5 Max.
Source: ByteDance
Combined with the experience, Xiaolei felt the confidence of ByteDance, that is, the level above the average line of the industry's large model + the processing cost is far lower than the industry. These two points are extremely attractive for any company that wants to develop a large model. At present, the bean bag model has gained a lot of partners on the B-side, and many companies from the automobile, mobile phone, PC and other industries have been connected to the large model service of Volcano Engine, including Geely Automobile, Celis, vivo, Xiaomi, Asus, etc. Driven by the cost-effective landing price, more and more enterprises will access the large model in the future.
At present, the development of large-scale model application is still in the early stage. According to QuestMobile data, as of March this year, the number of users in the AIGC industry based on large models was 73.8 million, an increase of 8 times year-on-year, accounting for only 6% of the number of mobile Internet users, and there is broad room for growth. ByteDance's highly competitive pricing is bound to create low-cost landing conditions for AIGC applications in addition to impacting the industry.
Backed by ByteDance's bean bag model, it's time to speed up the implementation of AIGC application scenarios.