Byte releases Doubao large model family, which can generate three copies of “Romance of the Three Kingdoms” for one dollar, and the price is 99% cheaper than the industry

What can one dollar buy?

At the 2024 Spring Volcano Engine Force Motive Force Conference held this morning, Volcano Engine President Tan Dai gave an unexpected answer.

One yuan can buy 1.25 million tokens of the main model of Doubao, which is about 2 million Chinese characters, equivalent to three books of "The Romance of the Three Kingdoms".

The price war begins! Byte releases large bean bag model

The key challenges in implementing large model scenarios lie in model effects, inference costs, and implementation difficulty.

In order to help enterprises solve these challenges and use better models, lower costs and easier-to-implement solutions to help enterprises achieve AI transformation, Volcano Engine officially launched a new generation of full-stack AI services today.

Good technology must be applied on a large scale and continuously polished with a large number of calls.

A year ago, the bean bag model formerly known as "Skylark" became one of the first large models in China to be registered through the algorithm.

After a year of iterative development, Doubao's large model currently processes 120 billion tokens of text and generates more than 30 million images per day.

Starting today, the Doubao model will be officially provided to the outside world through the Volcano Engine.

The model family released this time includes Doubao universal model pro, universal model lite, role-playing model, speech recognition model, speech synthesis model, Vincent diagram model and other models.

Among them, Beanbao Big Model Pro is the strongest representative of the beanbao model. It has excellent performance in multiple dimensions such as understanding generation logic and memory. It supports 128K context windows and can help users quickly understand long and difficult text content.

In some scenarios, for customers who are sensitive to delays and costs, the beanbag universal model lite is a better choice.

Thanks to the blessing of role-playing models, whether it is playing the role of a Sichuan dialect teacher or promoting the story interpretation of a script, the actual demonstration on the spot is very easy.

Voice is an important part of the interaction with AI.

Speech recognition models and speech synthesis models can accurately identify the user's content, language and context, and through learning of timbre, tone and intonation, large models can express real emotions, allowing AI to communicate like real people.

A large amount of usage can polish a good model and significantly reduce the unit cost of model inference. The Volcano Engine directly transformed into a "price butcher" today, bringing a billion-point shock to the large model market.

The pricing of models with the same specifications on the market is generally 0.12 yuan/thousand Tokens, while the Doubao universal model pro-32k model inference input price is only 0.0008 yuan/thousand Tokens, which is 99.3% lower than the industry price.

Or, the input price of the Doubao general model pro-128k model is 0.005 yuan/thousand Tokens, which is 95.8% lower than the industry price, truly achieving price involution "from cents to cents".

In addition, in order to help enterprises better implement large model applications, Tan Dai also announced the launch of the Volcano Ark 2.0 platform and released three large model plug-ins:

  • Networking plug-in: Search for relevant data information from data on the entire network
  • Content plug-in: Provide rich video and graphic content
  • Knowledge base plug-in: supports fine-tuning of proprietary data to reduce AI illusions as much as possible

Application-oriented, implementation is the way to go

Through an AI-native development platform, we have the opportunity to make everyone a developer of AI applications.

At the press conference, Tan Daidi made the above judgment eloquently. As a new generation AI application building platform, users can quickly build various model-based bots on Button, regardless of whether they have programming skills or not.

Moreover, users can also publish bots to various social platforms, communication software, or deploy them to other channels such as websites.

The opening of the press conference demonstrated the AI ​​bot of a fifth-grade child. It used the Chinese knowledge learned in school to create an AI bot that can be called an English foreign teacher.

In corporate environments that require higher application scenario capabilities, Haidilao uses buttons to simulate customer dialogue drills to help customer service ladies improve service levels.

Super Orangutan uses Kouzi to help users understand better fitness; China Merchants Bank built a mobile life discount bot based on Kouzi that recommends dining discount stores and a wealth perspective that analyzes market conditions…

The bot created by Button supports mixed interaction in multiple modes such as text, pictures, voice, video, GUI cards, etc. Users can even communicate with the bot by voice just like chatting with a real person, and it will always respond in the most intelligent and natural way. every request from the user.

In addition, Huoshan Engine has also officially released the Kouzi Professional Edition – according to the needs of enterprises, Huoshe Engine is further encapsulated based on the Kouzi platform and supports many advanced features.

In the past ten years or so, ByteDance, which understands the mobile Internet era best, always seems to be able to make a product that is popular, and has quietly become the behemoth it is today.

The most impressive thing about this conference is how ByteDance understands product application and development in the era of large models.

Zhu Jun, vice president of product and strategy, said that rather than running code on the server, it is more important to create appropriate product forms and natural interaction methods to meet the actual needs of users and make users truly willing to use these products.

At the press conference, Zhu Jun revealed the origin of the name of the bean bag.

"Doubao", which seems to have little to do with AI, is actually the first general principle that was determined when it was first named: simplicity, easy to read, and easy to remember.

At the same time, in order to shorten the distance between the product and the user, they have internally defined three product design principles for products such as bean bags: personification, proximity to the user, embedding in the user's usage environment, and personalization.

In the past year, ByteDance has done a lot of exploration in the form of large-model applications, and his biggest feeling is that compared with before the AI ​​era, there are both commonalities and big differences in making applications.

So what is the commonality? Human essential needs have not actually changed. For example, the need to obtain information quickly and conveniently has not changed, and we need to improve ourselves at work…

In his view, it was relatively easy to make products in the past, because at least the underlying technology was mature and stable, so you just had to use your empathy and think about what the users' needs were. But in the era of large models, it was completely different. different.

The new difficulty is not only to consider what the large model can do at this moment, but perhaps more importantly, to guess what new user scenarios can be implemented in 3 months, 6 months, or two years.

So this is a new challenge, and it is also necessary to continuously predict the PMF of the next product under the dynamic technological evolution.

Taking the AI ​​search engine as an example, Zhu Jun revealed that in the evaluation search task in the first half of last year, he often got 6 out of 10 questions wrong, which means that the search application scenario is completely untenable. However, with the evolution of model capabilities, now AI search tasks At least it is usable today.

This transformation from scratch, from usability to optimization, is not only a breakthrough at the technical level, but also the result of a deep insight into user needs.

According to a McKinsey report, by 2030, the global economic increment driven by large models will reach 49 trillion yuan, of which China's economic increment will reach 14 trillion yuan.

The huge economic increment includes not only the improvement of existing work efficiency by large models, but also the new scenarios and new business formats brought about by new technologies. ByteDance's exploration is a microcosm of the implementation of AI applications, and it is also a common topic that the entire industry needs to think about.

And this is exactly what Tan Dai repeatedly emphasized at this press conference. Good models must be used by everyone and every company.

# Welcome to follow the official WeChat public account of aifaner: aifaner (WeChat ID: ifanr). More exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo