The ChatGPT application with 'eyes' is here! Supported by a large multi-modal model with 15 billion parameters, it is versatile in reading pictures and chatting.

In many people’s minds, ChatGPT is an omnipotent “monster”.

It can write poetry, code, and even directly generate long papers... In addition, ChatGPT is also extremely emotional and expressive. It only takes a few rounds of conversations to conquer everyone.

But is ChatGPT really omnipotent?

As a chat robot, ChatGPT has excellent language understanding and text processing capabilities. However, due to the training model, ChatGPT does not support picture dialogue. Just imagine, when chatting with friends on weekdays, if you cannot send pictures, you will naturally lose part of the fun.

In China, there is such an AI R&D team. The "元成祥Chatlmg" application they launched can accurately identify pictures and chat with users, just like adding to ChatGPT A pair of "eyes", can not only understand text, but also perfectly identify pictures.

A few days ago, Magike.com learned from the Yuanchengxiang R&D team that "Yuanchengxiang Chatlmg" has just been launched recently. Follow the WeChat public account "Yuanchengxiang" You can try the application now. (Note: Limited by computing power, the initial access method is invitation-based)

So, what is the experience of using "Yuancheng Xiang Chatlmg" and whether its image recognition function is real What about magic? On the eve of the product launch, Magike.com conducted a comprehensive evaluation and summary.

Picture | Yuan Cheng Xiang Chatlmg Logo

Look at the picture and speak? This ChatGPT application is amazing

Enter "Yuancheng Xiang Chatlmg", you can see that the interface of the application is verysimple and clear-the text dialog box at the bottom and the symmetrical avatar layout , various elements are combined into our common chat software interface.

In the lower left corner of the application, you can see a conspicuous "Upload Picture" button.

Currently, the application supports three methods for uploading pictures, namely photo gallery, taking pictures and file import, one at a time.

Coinciding with Women’s Day, Meike.com also selected two items and handed them over to "Yuanchengxiang Chatlmg" to help determine whether they are suitable. One of them was from​​Search Engine​## The lipstick picture downloaded from #​, the other one is the notepad picture taken through the mobile phone:

It can be seen that "Yuancheng Xiang Chatlmg" accurately identified the products in the two pictures, and also extended some other descriptions, such as the bear on the notepad is cute and the lipstick is placed on a white background Wait a minute.

At the same time, he also gave his own judgment on whether these two items are suitable as gifts for girls, and also put forward rational suggestions whether the gifts are suitable based on the preferences of the recipient.

After "Yuan Cheng Xiang Chatlmg" gave the reply,

we asked another question:

Well, the suggestions are very pertinent, and the logic of contextual understanding is also coherent.

In addition, Meike.com also uploaded some other pictures for "Yuan Cheng Xiang Chatlmg" to identify.

For example, this one:

And for example, this one:

And these:

It can be seen that the recognition ability of "Yuancheng Xiang Chatlmg" is still quite accurate, and it candescribe what it sees in text form The picture content is .

Compared with ChatGPT based on pure text interaction, in addition to basic interactions such as writing poems and articles, "Yuancheng Xiang Chatlmg" with an extra pair of "eyes" has obviously discovered several new functions, such as item recognition. is one of its important functions.

After that,with the continuous expansion of data volume and continuous optimization of products, Meike.com also hopes that "Yuancheng Xiang Chatlmg" can bring us more outstanding performance.

Blessed by a large model with tens of billions of parameters, it will continue to be upgraded in the future

Essentially, ChatGPT is a chat robot product that provides the ability to generate text with the help of the GPT-3 large model . The power of ChatGPT is precisely due to GPT-3.

As a natural language processing model developed by OpenAI, it has too many talking points. Long before the launch of ChatGPT, GPT-3 had already successfully emerged in the AI ​​field. It is reported that the model reaches 175 billion parameters, and a single training requires tens of thousands of GPUs, costing nearly US$4.6 million. Such crazy “gold swallowing ability” can naturally support crazy applications like ChatGPT.

So are there any shortcomings in ChatGPT? The answer is yes!

According to the Yuanchengxiang R&D team, "Yuanchengxiang Chatlmg" is equipped with a self-developed

tens of billions of parameters multi-modal large model

, which is equivalent to adding "eyes" to ChatGPT , making it possible to talk more like a human being. The so-called multi-modality in the field of AI often refers to the collaboration of sensory information, such as

images, text, voice

, etc., to help artificial intelligence understand the external world more accurately. It is also with the support of multi-modal large models that "Yuancheng Xiang Chatlmg" has the ability to understand images.

According to the Yuanchengxiang R&D team, the current

Yuanchengxiang AI large model

parameter scale is about 15 billion; the data is currently mainly image-text data and VQA datawait. During the training process, the model releases as few learnable parameters as possible to ensure efficient training on limited computing power.

When asked about the model's future development path, Yuanchengxiang's R&D team revealed to Meike.com that in addition to the current interactive graphics and text, the large model will also use pictures in the future The ability to generate is combined, so that you can reply to your own generated images when answering, making the conversation more vivid.

In addition, this mode will also include more modes, such as video, audio, etc., and even be combined with digital human and other technologies to develop a more vivid three-dimensional image of.

It is conceivable that with the support of Yuanchengxiang’s multi-modal model with tens of billions of parameters, our conversations with chatbots will become more like real people, and the surprises that AI technology will bring us in the future will also more and more.

Chinese AI manufacturers are running wild, how far is the domestic ChatGPT?

At the "Ministerial Channel" of the First Session of the 14th National People's Congress held this month, ​​Minister of Science and Technology Wang Zhigang was talking about ChatGPT , praising the progress it has brought to AI and related industries.

At the same time, various discussions triggered by ChatGPT were also brought into the two sessions by representatives and committee members.

In fact, as early as 2017, before my country elevated the development of artificial intelligence to a national strategy, AI was already a "frequent guest" at the National Two Sessions. Now that ChatGPT is popular all over the world, people can't help but ask again: How can China's AI industry seize the trend of ChatGPT?

Taking advantage of the opportunity of new product release, Meike.com also had a discussion on this issue with Dr. Gao Yizhao

, head of Yuanchengxiang’s R&D team.

In his view, the construction of large AI models in China requires a long-term layout and continuity of work in order to develop truly innovative technologies

.

When we look back at the "history" of ChatGPT, we can find that it is not easy to build a large AI model. Behind it is long-term capital investment and iteration of computing power technology. In addition, we must also consider the training data Optimization, reduction of computing energy consumption and other aspects.

In other words,

If you want to make a ChatGPT-like application, you can not complete it through a simple "100-meter sprint", but you need to be patient and complete a "10,000-meter long-distance race". Only then can you enter the "finals" where the heroes compete.

.

However, judging from the current development status, the large AI models of domestic manufacturers are still in the research and development stage, and teams that can reach the application level like "Yuancheng Xiang Chatlmg" are still very rare - first of all, domestic The research and development of large AI models generally started late, and there is a serious talent gap. Secondly, the training of large models in the Chinese context is more difficult and takes far more time than English.

These problems should have required "patience" to solve, but as the last wave of AI investment boom cooled down, many companies were forced to reduce their budgets, which ultimately led to the development of related AI businesses reaching a bottleneck.

In contrast, the Yuanchengxiang R&D team has gone relatively smoothly. On the one hand, the team has long-term exploration in the field of AI applications, and on the other hand, it also benefits from their multi-modal pre-training. The correct judgment and foresight in the model, and finally with the persistence of the entire team, finally ushered in the " Blooming and Fruiting

" in the ChatGPT era.

When talking about the future exploration of AI companies, Dr. Gao also gave some of his own opinions: "

In terms of ecological construction, China can promote open source​ After all, there is strength in numbers.”

ConclusionSince its release last year, ChatGPT has evolved from a “phenomenal product” to a "

Technology Symbol

", major domestic technology manufacturers are rushing to launch ChatGPT-like products, hoping to win the so-called "victory" in this competition.

At the same time, with the popularity of ChatGPT, the domestic AI industry once again set off a new round of new product boom. It was during this period that "Yuancheng Xiang Chatlmg" stood out. Judging from the evaluation of the "Yuancheng Xiang Chatlmg" application, its experience is already very friendly, and it even has "image recognition" functions that ChatGPT does not have. Of course, as an early version of "Yuanchengxiang Chatlmg", there are still some deficiencies in some details, but I believe that with the efforts and persistence of the Yuanchengxiang R&D team, the domestic ChatGPT application can also surprise the industry in the future.

The above is the detailed content of The ChatGPT application with 'eyes' is here! Supported by a large multi-modal model with 15 billion parameters, it is versatile in reading pictures and chatting.. For more information, please follow other related articles on the PHP Chinese website!