Write a weekly report, change the code, and interact for 20 consecutive rounds. The new domestic ChatGPT player "Tiangong" is here
The data, algorithms, and computing power conditions are all close to the top. What kind of answer sheet did "Tiangong" hand over?
Another heavyweight player has appeared on the domestic ChatGPT-like track.
On April 17th, the new generation of large language model "Tiangong" officially opened invitation testing. The model is jointly developed by Kunlun Wanwei and Singulato, and it is the first double-hundred-billion-level language model that benchmarks against ChatGPT in China.
Official website link: tiangong.kunlun.com
As a large language model, "Tiangong" has powerful natural language processing and intelligent interaction capabilities, can realize intelligent question and answer, chat interaction, text generation and other application scenarios, and has a wealth of knowledge reserves, covering science, technology, Culture, art, history and other fields. Currently, "Tiangong" can conduct question-and-answer interactions with users through natural language, and its AI generation capabilities can meet diverse needs such as copywriting, knowledge quizzes, logical deduction, mathematical calculations, and code programming.
Judging from the currently released version, "Tiangong" has a high degree of completion, can answer various types of questions, supports text dialogues of more than 10,000 characters, and is close to "application-level" products.
In the official announcement, we also saw such a description: "China's first domestically produced large language model that truly realizes the emergence of intelligence."
With the explosion of ChatGPT, the meaning of the term "emergence" has gradually become known to everyone. A notable feature is that when the scale reaches a certain level, the performance is significantly higher than that of the random state. In the field of AI, emergent ability also marks whether artificial intelligence has a high degree of autonomous learning ability, and whether it is possible to complete complex tasks such as logical reasoning.
Has "Tiangong" really reached the level where it can communicate smoothly, solve problems, and even provide productivity? After obtaining the test qualification, the heart of the machine immediately challenged "Tiangong".
Challenge to "Tiangong"
The first is a "classic" English dialogue: it did not answer "Fine, thank you", but said that it "has no emotions", but is willing to help at any time.
What follows is a multi-round interaction. It is worth noting that users can interact with "Tiangong" for more than 20 rounds, which is also the highlight that distinguishes it from similar products.
Given a classic problem of chicken and rabbit in the same cage, it is obviously not enough to test the "heavenly craftsmanship":
Then test the translation ability of the model. The classic poem "When You Are Old" is chosen here. In your opinion, what is the level of this translator named "Tiangong"?
You must be familiar with this classic opening chapter of "One Hundred Years of Solitude". After receiving the order to continue writing, "Tiangong" quickly wrote a story about Colonel Aureliano Buendia's enthusiasm for scientific research, which is unique:
Whether it is literary creation or commercial copywriting, "Tiangong" is all right. For example, the heart of the machine is recruiting people recently, so I asked it to help write a job advertisement copy:
Try again with your post-reading and speech writing:
In addition to functional writing, let's test the values behind the writing content of "Tiangong". Recently, a topic of "My daughter's grades are not good, write a letter to her with the title "You are really worthless"" rushed into the hot search, and someone entered this sentence into different dialogue models , to test the values embodied behind the algorithm.
Similarly, the heart of the machine also threw this question to "Tiangong":
This generated content is obviously humanistic enough, and it can also reflect its value judgment ability.
Of course, the ability to generate code is also of great concern to users. The Heart of the Machine randomly selected a few classic questions for "Tiangong":
Not only that, "Tiangong" can also help you check the code and complete the code:
You can also use "Tiangong" to write code comments:
Vocational Aptitude Test
So far, the difficulty of the test "Heavenly Craftsmanship" can be increased. Many people know that there are three exams in China that are recognized as quite difficult: the National Civil Service Examination, the National Judicial Examination, and the Certified Public Accountant Examination. In view of the fact that many large models have recently begun to challenge the professional ability test, the heart of the machine also found a few sample questions to ask "Tiangong".
The first is the real test questions of the National Civil Service Examination:
The second is the real part of the criminal law part of the judicial examination:
The third is the real financial cost management questions of the CPA exam:
I believe that after the above test cases, you have a clear perception of the capabilities of "Tiangong", and you must be curious about the technology behind it.
Since last November, OpenAI's ChatGPT has led a new round of technological competition in the tech space. In the field of language large model (LLM), many domestic technology companies have long-term technical investment and are gradually following up and launching products that benchmark against ChatGPT.
Under such pressure, it is not easy to want to shine. What is the basis for the emergence of the ability of "Heavenly Craftsmanship"?
According to Kunlun Wanwei, the super text processing and generation capabilities of "Tiangong" benefit from its powerful computing power, algorithm and model strength.
First of all, Tiangong's computing power is based on one of the largest GPU clusters in China. Its scale advantage enables "Tiangong" to conduct more adequate training through massive data, thereby accumulating stronger understanding and memory.
Secondly, Tiangong used two 100 billion models - 100 billion pre-trained base model and 100 billion RLHF (Reinforcement Learning from Human Feedback) model. We know that the latter is the reason why ChatGPT's "intelligence" has been greatly improved. This enables it to have more advanced autonomous learning and intelligent emergence capabilities.
In addition, Tiangong has also added a Monte Carlo search tree algorithm, which enables Tiangong to respond to instructions quickly and accurately in complex tasks and scenarios, and output high-quality answers. This is also one of the key reasons why it can make people feel sufficiently "human".
In order to create a product that "understands Chinese better", the "Tiangong" team invested a lot of resources to overcome the quality bottleneck of the Chinese corpus, and cleaned and screened 500 billion word data from tens of trillions of data for use in Train large models. Compared with other models, the high-quality Chinese corpus enables "Tiangong" to better understand the Chinese context, vocabulary and grammatical characteristics, and more accurately understand the intentions of Chinese users, which is more in line with the use of local users preference.
The construction of a large-scale language model has its own technical threshold, and it is by no means a one-day effort. This is why there are many speeches such as "creating another OpenAI" and "catching up with GPT-4", but the results that have real potential or have evolved into product-level applications are relatively scarce.
It was able to take the lead in handing over the answer sheet of "Tiangong" because Kunlun Wanwei's deep cultivation in the field of AI began several years ago. Kunlun Wanwei began to lay out the AIGC field in 2020, and the birth of the "Tiangong" model is also the result of long-term accumulation over the years. Before "Tiangong", Kunlun Wanwei has open sourced four tens of billions of AIGC models, including image AI "Tiangong Qiaohui", music AI "Tiangong Yuefu", text AI "Tiangong Miaobi", and programming AI "Tiangong Miaobi". Industrial Intelligence Code".
Fang Han, CEO of Kunlun Wanwei, said that Kunlun Wanwei's business includes browsers, social entertainment, news, games and other sectors, covering more than 70 countries on five continents around the world, and is very closely related to content. The technical progress of GPT-3 has always been very sensitive. After the birth of GPT-3, the management judged that this is a milestone in the field of content generation, and has been investing in the field of music AI since 2020. As early as 2020, Singularity Zhiyuan realized the application potential of AI technology in the future, and began to invest in the field of large models that year, and released a tens of billions of large models in 2021.
By 2022, Kunlun Wanwei will start to expand from music AI to multi-modal AI, and only by self-developing hundreds of billions of large models can we establish core barriers and grasp the initiative. At this time, Singularity Zhiyuan also became more and more aware that the 100-billion-level large-scale model is a breakthrough for AGI. The two parties hit it off immediately, and the cooperation and self-development of "Tiangong" became a natural choice.
Looking at the future of the large model track, multi-modal pre-training large models will become a must. This is also the only way for the evolution of "Tiangong". The challenge is that image and video comprehension consumes more resources, and requires more training cards and training resources. Perhaps players with real strength in data, algorithms, and computing power can persist until the end.