Author | Yu Jinglin
Edit | Tang Yongyi
Bytes, which has been silent for half a year, has finally released the latest progress of its own large model!
Without a little signal, on August 17, ByteDance announced that it began to test the AI dialogue product "Beubao", enter the official website (https://www.doubao.com/) to directly register and use or download the App through the QR code provided by the official website.
Source: Doubao app
Before the release of "Beubao", Byte's latest AI progress was the internal AI project "Grace", which was exposed in June.
According to Tech Planet, Grace began a small beta in July, and now changes its name to "Bean Bao" and officially meets the public.
However, according to the latest information obtained by 36Kr, Byte Grace can not be regarded as the predecessor of Bean Bao, Grace is positioned more like an AI application shelf internally, and "Bean Bao" is only the first product developed by Grace's team. At present, scan the App download button on the official website of Doubao, and the official website will show that Android and IOS Apps can only download the beta version.
How is the "bean bag" experience?
From the name "Bean Bao" and the cute virtual 3D image, it can be seen that Byte does not want to make a "Xueba" tool like ChatGPT, but a more people-friendly product.
Entering the default page, you can find that "Doubao" divides the page into two parts, and on the left are four AI robots, including Doubao, Xiaoning who loves to chat, English learning assistant and all-round writing assistant. On the right is the dialog box corresponding to the AI robot.
Source: Doubao app
What is the difference between the four AI assistants?
Listen to the doubao self-introduction——
Source: Doubao app
To sum up, the four of them have their own strengths -
Bean bun is the squad leader, where it needs to be moved;
Xiao Ning, who loves to chat, is a good table mate in China, and chatting is funny and a good partner;
The English learning assistant is the representative of English class, improving your English score in all aspects;
The all-round writing assistant is the representative of the language class, and literary literacy depends on it.
But that doesn't mean users can only use these four default AI bots, just as you can create a new dialog on ChatGPT, you can "start a new one" by creating a new dialog, and you can edit the name of the new dialog or delete the conversation.
Source: Bean Bao
If you ask Doubao: Who is Doubao? What can be done?
Source: Bean Bao
Source: Bean Bao
Since the birth of ChatGPT, what are the advantages and disadvantages of ChatGPT compared with generative artificial intelligence, Beubao and other generative artificial intelligence big language models?
Source: Bean Bao
Let's take a look at the "language understanding ability" of "Bean Bao".
Source: Bean Bao
If asked about the four famous books, "Bean Bao" answered accurately and performed well in understanding the content of the novel.
In contrast, Bean Bao is quite honest in "summarizing" the disadvantages - its accuracy really leaves something to be desired.
When we randomly enter a math problem:
Source: Bean Bao
The correct answer should be [0,1], and even if you answer again, you still provide the wrong answer...
Source: Bean Bao
What if it wasn't a math subject?
Source: Bean Bao
Answer correctly!
In addition, judging from the evaluation, "Bean Bao" only supports answering "scientifically based" questions.
Source: Bean Bao
In addition to Chinese, "Bean Bao" also supports English conversations, click the upper right corner to switch languages.
Source: Bean Bao
There are rumors that the domestic large model lacks its own database, and needs to be trained with a lot of data from English-speaking countries, whether "bean bag" will be like this, switch the language to test:
Judging from the jokes, the performance of "bean bag" in Chinese and English is half a pound and eight taels.
Source: Bean Bao
When you are bored, you can chat with "Xiaoning who loves to chat", is it a qualified chat partner? Let me test it!
Source: Bean Bao
Xiao Ning's "default state" is like a capybara, quite stable and serious. But it's very obedient, and it switches styles when you want it to switch styles.
The all-round writing assistant not only generates poems on request, but also provides reading appreciation.
Source: Bean Bao
Although the verse of this "Song of the Field" is not so beautiful, the poem only took less than 10 seconds.
When we asked the English Learning Assistant a question, how did it perform?
Source: Bean Bao
I tried to ask it about the difference between the words "ban" and "interdict", and I could see that the answer of the "English learning assistant" was not entirely in English.
If you feel that the English answer is still difficult to understand, you can let the "English Learning Assistant" convert it into a Chinese, bilingual switching, very silky ~
Overall, "Bean Bao" can currently support some simple questions and answers, and the accuracy needs to be improved.
Moreover, now Doubao only supports text generation, and there is no image function for the time being.
Bean Bao is also not yet online. According to Byte's official introduction, the training data of "Bean Bao" is as of 2023, so it is impossible to answer relevant questions in 2023.
Burying his head in hard work for half a year, the byte large model "sneaks"
Compared with other domestic Internet companies, although Byte began to form a large-model team as early as February, it was long overdue to the real release of the first product "Bean Bao".
Byte has been deployed in the field of artificial intelligence for a long time. In 2016, Byte established AI Lab, an artificial intelligence laboratory, focusing on natural language processing, machine learning, data mining and other aspects of research.
However, after the domestic "big model war" set off, Byte has not announced the progress of any underlying large model.
On the contrary, in June, Volcano Engine first released the large model service platform "Volcano Ark" to provide computing power and other supporting services to domestic large model companies.
Source: Volcano Engine
And until the recent release of bean bags, the Byte Large Model team has actually quietly released a lot of progress.
In July, a large model called BuboGPT was officially launched on Github, also backed by the ByteDance team.
BuboGPT is a multimodal large model - BuboGPT can integrate multimodal inputs such as text, image and audio, and shows excellent dialogue capabilities in image and audio data processing.
Official Website:https://bubo-gpt.github.io/
Source: BuboGPT
The best thing about BuboGPT is listening to the position.
Source: BuboGPT
For example, you only need to upload an image of audio, and BuboGPT can quickly identify where the audio comes from the image, and intimately use the bounding box to help you frame it!
From BuboGPT, you can also see the direction of the byte-large model - combine images and text, and try to apply multi-modality.
It can be imagined that this is closely related to Byte's business - as an APP factory, Byte has a number of C-end applications such as Douyin, and its application scenarios in audio and video are very rich, and it is reasonable to make efforts in this regard.
Kuaishou AI to "kick the pavilion"
However, the bean bag thrown by Byte is still hot, and Byte's old rival, Kuaishou, obviously can't hold it.
On the evening of August 18, Kuaishou suddenly announced that the Android version of "Kuaishou AI Dialogue" was open for internal testing - after that, Kuaishou AI had already conducted a wave of small-scale testing in early August.
But Byte Bean Bao is still an independent app, but obviously Kuaishou wants to take the lead in the app, and Kuaishou AI dialogue is directly built into Kuaishou.
In terms of usage, users only need to click the AI icon in the upper right corner of Kuaishou Search homepage, directly enter the content or click the guide question to start the conversation.
Alternatively, users can trigger the "Search Smart Q&A" card to use this feature through the bottom entry.
Source: Kuaishou
With the support of large models, users can find the anchors or videos they are interested in more efficiently.
Source: Kuaishou
Source: Kuaishou
Kuaishou AI Dialogue is also the first application product based on large language models in the short video live broadcast industry.
The domestic large-model melee is still hot, and when manufacturers are working the underlying large-language model, the competition on the application side has also begun——
AI large models have gradually entered various terminals and national applications. For example, recently, Huawei and Xiaomi have put large models into their mobile phone systems.
Nowadays, "jitter fast" has released AI applications, which is undoubtedly another important node for the landing of domestic large models.
Source: Open Web
Resources:
https://arxiv.org/pdf/2307.08581.pdf