HomeAI News
Domestic ChatGPT "Serial Monkey" first test on the whole network! No need to wait for online experience, drawing Soviet Musk is super realistic

Domestic ChatGPT "Serial Monkey" first test on the whole network! No need to wait for online experience, drawing Soviet Musk is super realistic

Hayo News
Hayo News
April 24th, 2023
View OriginalTranslated by Google

In the ChatGPT domestic competition, another player submitted the paper. In February of this year, the domestic competition officially declared war, and various forces took turns to play. At that time, Li Zhifei, the founder and CEO of Wenwen, was one of the high-profile entrepreneurs. After just over two months, the results came out.

On April 20, the self-developed large-scale model "Serial Monkey" officially debuted on Wenwen. Based on this large model, Mobvoy also launched a one-stop CoPilot AIGC product matrix for creators. Specifically, it includes the following four products: AI writing platform "Wonderful Text", AI painting platform "Yanzhihua", AI dubbing platform "Magic Sound Workshop", and digital human video and live broadcast platform "Wonderful Yuan". Key points: Four products are open for registration, all of which are directly available!

Fantastic text

It is understood that the four AIGC products of Going out to ask are all open for experience on the day of the press conference, and the voice cloning function of "Magic Sound Workshop" will be officially launched the day after tomorrow. What are you waiting for, hurry up and rush first.

Experience address: write.mobvoi.com

Open the homepage, and there is a big sign in the upper right corner: Login/Register. We only need to register an account with our mobile phone number, and we can start to experience it, which is completely different from other products that require an invitation to conduct internal testing. Moreover, the amount given at the beginning seems to be quite sincere.

According to the official introduction, the AI writing assistant "Wonderful Text" covers the four major content creation scenarios of workplace office, marketing, new media and creative writing. And 8 major AI editing functions are opened, namely style conversion, key point extraction, proofreading and error correction, continuation, rewriting, expansion, abbreviation, and translation. By the way, we can use the following box in "Wonderful Text" to directly invoke the language, knowledge, logic, mathematics, reasoning and problem-solving planning abilities of "Sequence Monkey". It can easily implement functions such as continuous dialogue, and even let it match a piece of text with pictures.

It is equivalent to the new media copywriting, writing, videos, Xiaohongshu, Zhihu "Wonderful Articles" are all included.

AI, give me a job

The test starts, first let it write a profile for the "Xinzhiyuan" account. In the prompt box, enter the account name, target users, specific description characteristics, and key information.

The first paragraph of the output is quite satisfactory, covering a few points of the prompt.

What is unexpected is that in the next generation of content, we not only point out that we provide cutting-edge technology information, but also provide an interactive platform. Well, you know "New Wisdom".

Let "Wonderful Text" title "SpaceX Starship First Flight Explosion".

Give a specific content: After experiencing delays, Elon Musk’s space exploration technology company SpaceX’s new generation of launch vehicle Starship finally ignited and launched, but it exploded in the air, and the first orbital test flight ended in failure. Musk then posted a response, first congratulating the SpaceX team for their efforts, and saying that he learned a lot for the launch again a few months later. "Wonderful Article" is concise and concise, giving the title "SpaceX starship test flight failed, Musk: I learned a lot!". After choosing "Free Humor" in the style change, it began to refer to Musk affectionately as "Academician Ma".

Moreover, it will also extract key points from the text and start titles one by one.

Generate a map, that's it

Zibo BBQ, which has recently gone viral, has become a national-level Internet celebrity check-in place.

The most attractive thing about Zibo barbecue is its unique "eating method" of small skewers + small cakes + shallots, and a three-piece set of small cake oven plus dipping sauce and soul barbecue was born. It is better to use "wonderful text" to write a grass copy in Xiaohongshu.

Not only did I use a variety of emoticons, but I also analyzed everything in just a few lines of words, and finally attached the hashtag #, which is really interesting. Although the three-paragraph writing method is a bit of a drama, the pictures generated by AI based on this content are still very amazing. I have to say that these two skewers of grilled meat are really mouth-watering.

And after letting "Wonderful Text" generate a self-introduction, we also let it match a photo for ourselves.

It's okay, there is no sense of disobedience, very technical man (dog head). It turns out that in the heart of this model, I look like this.

Open a brain hole: car show + ice cream

Recently, the topic of "Auto Show Ice Cream" has become quite popular. Why not let "Wonderful Article" tell us a talk show with the title "Selling Ice Cream at the Auto Show".

Don't tell me, it's quite creative. In order to attract the wealthy, it proposed to launch ice cream in "Ferrari red", "Porsche yellow" and "Lamborghini green" flavors.

In order to attract car fans, it thinks it can put some wheels and door handles on the ice cream.

Well, it's making a very new ice cream.

In addition, "Wonderful Text" can also provide us with short video inspiration. Then ask it to provide us with the idea of the short video "Long Aotian swore to protect Liu Bo".

Unexpectedly, "Wonderful Essay" is in the style of chicken soup, "True courage is not not being afraid, but being able to protect the one you love when you are afraid."

The editor admits: Damn, I was moved by this AI.

How about doing math and writing code?

Now, whenever a large language model is evaluated, everyone will debut "Chicken and Rabbit in the Same Cage" and give it a try. Seeing this posture, various R&D teams probably also trained the models individually overnight. In order to avoid the hidden mechanism that may be triggered by the words "chicken and rabbit in the same cage", we specially replaced it with "radish and cabbage" in the test: Xiao Ming and his mother went to the market to buy vegetables. It is known that cabbage is 5 yuan per catty, and radish 8 yuan per catty, I bought 10 catties of cabbage and radish in total, and spent 62 yuan, so how many catties of cabbage and radish did I buy? Yes, the first time successfully solved the answer. Given the random nature of the model, we ran several more tests. But the result is that sometimes you can do it right, and sometimes you can do it wrong, and the effect is relatively average.

In addition, "Wonderful Text" can actually handle some simple codes easily. But it may be a problem with the document settings. After adding the generated code, the format is gone... For example, let it write a piece of Python code to calculate factorial. Obviously, this indentation is wrong.

Now, let's let "Wonderful Text" modify it. Well, the AI not only judged that there was nothing wrong with the code itself, but also successfully modified the indentation.

But after adding it to the document, it broke in a second...

A writing assistant

However, judging from the official introduction, "Wonderful Text" currently focuses on the text editing scene. "Wonderful Text" can provide users with a steady stream of inspiration and creative direction when writing year-end summaries, customer service speech skills, scripts, advertising copywriting and other content. At this point, according to the experience just now, usually after a few more attempts, there will be basically satisfactory results. As for problems in science and engineering such as mathematics and code, the model is still lacking, but it is said that the team is already improving.

painting of words

Next is another project that has been open for experience-AI painting.

Experience address: paint.mobvoi.com

Also, sign up to get 5 trial opportunities. In order to make up the number of evaluations, the small partners in the editorial department resorted to their prehistoric powers and pulled colleagues from the whole company to register one.

It takes 10 points to generate a picture. When using it, "Painting of Words" provides 8 styles of pictures, but you can choose not to.

However, the type of model is mandatory, such as two-dimensional, general, photo, etc.

First, something simple: a glass of orange juice. This picture generated by "Painting of Words" can be scored as a passing grade. The changes of light and shadow, and the reflection of the glass are basically consistent.

Now, let’s change to a more complicated one: “Musk works feverishly in a Soviet factory, and his life is thriving.” I believe many people have discovered that when using AI to draw tools, they suddenly become “dead language”. The intelligent text generation function of "Yanzhihua" is here to save us. We only need to fine-tune the content generated by "Yanzhihua", and we can get more perfect prompts:

According to this prompt, "Yanzhihua" generated this picture. And what I didn't expect was that Academician Ma's hands were handled quite well.

Realistic character models However, these rough tips are obviously not amazing enough, why don't we ask for help from other people's professional "spells":

female ornate princess, with white long flowing hair, bright beautiful eyes, trending on artstation, flowers of hope by Jean-Honor Fragonard, Peter mohrbacher, hyper detailed, insane details, stunning, intricate, elite, art nouveau, ornate, liquid wax, elegant, luxury, Greg Rutkowski, ink style, sticker, vector-art beautiful character design, double exposure shot, luminous design, award winning, masterpiece, amoled black background

Try it first with the "General Model". The elegant princess has long white hair and big and bright eyes staring into the distance. It can be said to be quite gorgeous.

Look at the "realistic model" again, the beauty is natural, it is simply overwhelming.

It is said that some prompt masters have already used the "Painting of Words" to generate the following divine pictures.

The applet is online!

In addition to the traditional web version of Wen Shengtu, the mini-program version of "Yanzhihua" is also coming these two days! Not only can you customize various styles of AI avatars, but you can also generate animation videos with one click.

With the few experience times left, the editor decisively chose the "magic likeness" that can change the style of photos.

Editor No. 1:

light and shadow

Fengchao Xiaobian No. 2:

Xianfeng Xiaobian No. 3:

light and shadow

Magic Sound Workshop

In addition to "Wonderful Text" and "Words of Words", there are two most mature products on Wenmenbang: one is the industry's top-level AI dubbing platform, and "Magic Sound Workshop" is used by tens of millions of fans and big Vs. ", and the other is the AI digital avatar "Wonderful Yuan", which can do digital human video and live broadcast. With the blessing of the "Sequence Monkey" large model, "Magic Sound Workshop" (overseas version DupDub) is the world's first dubbing platform equipped with a large model AI writing function, covering multiple scenarios such as AI writing, AI dubbing and editing. Here, you can choose from thousands of AI timbres, over 2,000 voice styles, 40 languages and 11 dialects. Easily complete content creation that integrates copywriting and dubbing, such as film and television commentary, audiobooks, online education, and news broadcasts.

Experience address: moyin.com In addition, "Magic Sound Workshop" supports the adjustment of 7 emotions including calm, sad, and happy for the selected voice, and 10 kinds of characters including middle-aged women and boys. migrate. At the same time, AI voice personalization editing functions such as rhythm adjustment, local speed change, and multi-person dubbing are also available, allowing users to edit voices like editing documents with word. In addition to selecting and editing voices, "Magic Sound Workshop" will also launch the "Pinch Voice" function, so that you can freely select voice characteristics such as gender, age, language, style and emotion, and create your favorite voice from 0 to 1 up.

Fantastic element

The dub is there, give it the perfect look. The AI digital image creation and live broadcast platform "Wonderful Yuan" can do exactly this. Currently, "Wonderful Yuan" has over 100 digital humans, over 1,000 3D digital assets, and over 1,000 voices.

Experience address: weta365.com With multi-modal generation technology, "Wonderful Element" currently supports three different types of image modeling (2D digital human), video modeling (2.5D digital human), and 3D modeling (3D digital human). forms of digital human generation. Its image cloning function only needs a 5-minute real-life video material to reproduce the user's image and demeanor 1:1, creating a digital human clone with the same voice and natural movements for the user.

Why is it called Sequence Monkey?

Compared with the 175 billion parameters of GPT-3, the number of parameters of "Sequence Monkey" is about tens of billions, which is a multi-modal large language model. The reason why it is named "serial monkey" is because its inspiration comes from the "infinite monkey" theorem. This is a famous thought experiment proposed by the mathematician Emil Borrel in the early 20th century. According to the theorem, if a group of monkeys hit a typewriter at random, the complete works of Shakespeare must eventually be produced. This theorem, based on probability theory and combinatorics, illustrates the unity of probability.

Going out to ask about the construction principle of the self-developed large model is similar to "Infinite Monkey". Massive text sequences are continuously trained by autonomous algorithms, digested and understood by large-scale computing power, and become "sequence monkeys". At present, this large model already has certain natural language understanding, knowledge, logic and reasoning capabilities.

All In AGI: The Road to Large-scale Models of "Going Out and Asking"

It can be said that with the blessing of the "sequence monkey" large model, the CoPilot ability specially created for creators is really impressive. So, why is this artificial intelligence company able to take the lead in the domestic large-scale entrepreneurial track? According to public information, Li Zhifei, the founder of Momenwenwen, once studied for a doctorate at Johns Hopkins University in the United States, and has been researching natural language processing and machine translation. After graduation, he joined Google and led the development of a series of products such as offline translation of Google mobile version. In 2012, Li Zhifei, a former Google AI scientist, returned to China after resigning. He summoned engineers from Stanford, Maryland, MIT, Tsinghua University, Peking University and other famous universities at home and abroad to form a group called "American Famous School Collector". To 1 has built core technologies such as speech recognition, semantic understanding, dialogue management, and vertical search, and has always maintained a world-class level.

In 2015, when Li Zhifei was seeking the commercialization path of AI voice, he chose the smart hardware track and released a series of star-level smart watches. In the next 2-3 years, Mobwang released several different smart hardware products. In this process, the company has accumulated a complete voice interaction technology stack and the ability to combine software and hardware. In 2020, Mobvoy began to deploy generative AI, and GPT-3 was born, which ignited Li Zhifei's enthusiasm for AGI. Subsequently, he formed a team of more than a dozen people, and released the large model UCLAI in 2021. At the same time, it also launched the first AIGC commercial product - the AI dubbing platform "Magic Voice Workshop". This platform has received over one million paying users around the world, officially opening the prelude to the commercialization of AIGC.

Immediately afterwards, Mobvoi continued to develop and deploy AIGC's underlying modular technology, including Wensheng voice, Wensheng pictures, and Tusheng video. In 2022, the company has achieved the full commercialization of the AIGC product matrix. After ChatGPT became popular, Li Zhifei was once again excited. He flew to Silicon Valley of the United States twice to exchange ideas with engineers from OpenAI, Google, DeepMind and other major factories. After many internal iterations, until today, Mobask has released the amazing "Serial Monkey" model. The ability system of "Sequence Monkey" with language as the core covers six dimensions of "knowledge, dialogue, mathematics, logic, reasoning, and planning", and can simultaneously support different tasks such as text generation, image generation, 3D content generation, speech generation, and speech recognition. .

In addition to AIGC products for creators and Magic Questions for C-end users, on the B-end, "Sequence Monkey" will open up its ecology to empower more industries and allow more companies to create their own large models. CoPilot. By providing API services to users, the large model "Sequence Monkey" can not only provide general ability support services and digital human image customization services, but also open up its own role capabilities and continuously iterate data interfaces. Enterprise users can also train on industry-specific content by uploading documents, so as to realize customized voice interaction.

Open platform address: openapi.mobvoi.com

It is understood that Momenwenwen has reached cooperation with the first batch of internal testing and exploration partners in the top ten industries, including automobiles, education, law firms, finance, medical care, tourism and other fields. Since its inception, Mobwenwen has always been a pioneer of domestic language technology, and it is a company that really does technology. In the past, its AI capabilities were implemented in the form of wearable hardware. In the future, it will empower enterprises and individuals in the form of CoPilot. It can be said that the ability of the "sequence monkey" model we have seen is in line with the AI technology accumulation of more than ten years when I went out and asked myself. From voice assistants to intelligent assistants, it is the result of ten years of hard work in the field of AI. . As Li Zhifei said, "Serial Monkey" is still a baby who is improving every day. Through continuous learning, I believe in the future, the "sequence monkey" model will bring us more surprises.




Reprinted from 新智元View Original


no dataCoffee time! Feel free to comment