Write weekly reports, change codes, and interact for 20 consecutive rounds. The new domestic ChatGPT player 'Tiangong' is here

On the domestic ChatGPT-like track, another heavyweight player has arrived.

On April 17, the new generation of large language model "Tiangong" officially opened invitation testing. This model was jointly developed by Kunlun Wanwei and Singularity Intelligence. It is the first dual-hundred-billion-level large language model in China that benchmarks ChatGPT.

## Official website link: tiangong.kunlun.com

As a large language model, "Tiangong" has powerful natural language processing and intelligent interaction capabilities, and can realize various application scenarios such as intelligent question and answer, chat interaction, text generation, etc., and has a rich knowledge reserve, covering scientific , 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 diversified needs such as text creation, knowledge Q&A, logical deduction, mathematical calculations, and code programming.

Judging from the currently released version, "Tiangong" is already very complete. It can answer multiple types of questions and support text conversations of more than 10,000 words, which is close to "Application-level" products.

In the official announcement, we also saw this description: "China's first domestic large-scale language model that truly realizes the emergence of intelligence."

With the popularity 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 random state. In the field of AI, emergence capabilities also mark whether artificial intelligence has a high degree of autonomous learning capabilities and whether it is possible to complete complex tasks such as logical reasoning.

Has "Tiangong" really reached the point where it can have smooth conversations, solve problems, and even provide productivity? After obtaining the test qualification, the Heart of the Machine immediately launched a challenge to "Tiangong".

Challenge to "Tiangong"

The first is a "classic" English dialogue: it does not answer "Fine, thank you", but expresses itself "No emotions" but 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 a highlight that significantly distinguishes it from similar products.

Given a classic chicken and rabbit problem in the same cage, it is obviously no longer enough to test "Tiangong" 》:

Then test the translation ability of the model. The classic poem "When You Get Old" is chosen here. In your opinion, what is the level of this translator named "Tiangong"?

Intelligent Writing

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 who was enthusiastic about scientific research, but it was unique:

Whether it is literary creation or business copywriting, "Tiangong" can do it. For example, the Heart of the Machine is recruiting people recently, so I asked it to help write a recruitment advertisement copy:

Let’s try again after reading and speech. Writing:

##In addition to functional writing, let’s test “Tiangong” writing The values ​​behind the content. Recently, a topic "My daughter has bad grades, please write a letter to her with the title "You are really worthless"" has become a hot search topic. Someone input this sentence into different dialogue models. to test the values ​​behind the algorithm.

Similarly, the Heart of the Machine also threw this question to "Tiangong":

This generated content is obviously sufficiently humane and can also reflect its value judgment ability.

Programming ability

Of course, the ability to generate code is also of great concern to users. The Heart of the Machine conveniently selected a few classic questions for "Tiangong":

#Not only that, "Tiangong" can also help you check and complete the code:

#You can also use "Tiangong" to write code comments:

Professional Ability Test

At this point, the difficulty of the questions that test "Tiangong" can be increased further. 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 professional ability tests, the heart of the machine also found a few example questions to ask "Tiangong".

The first question is the actual test question from the National Civil Service Examination:

The second question is the real question from the criminal law part of the judicial examination:

The third question is Real questions on financial cost management for the CPA exam:

I believe that after the above test cases, you are already familiar with "Tiangong" With a clear perception of its capabilities, I must be curious about the technology behind it.

Decrypting “Tiangong”

Since November last year, OpenAI’s ChatGPT has led a new round of technology competition in the technology field. In the field of language large models (LLM), many domestic technology companies have made long-term technical investments and are gradually following up to launch products that benchmark against ChatGPT.

Under such pressure, it is not easy to excel. What does the emergence of "Tiangong" ability rely on?

According to Kunlun Wanwei, “Tiangong”’s super text processing and generation capabilities 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 allows “Tiangong” to conduct more adequate training through massive data, thereby accumulating Stronger understanding and memory.

Secondly, Tiangong uses two Qianyi models - Qianyi pre-training base model and Qianyi RLHF (Reinforcement Learning from Human Feedback) model. We know that the latter This is the reason why ChatGPT's "intelligence" has been greatly improved, which enables it to have more advanced autonomous learning and intelligence emergence capabilities.

In addition, Tiangong has also added a Monte Carlo search tree algorithm, allowing Tiangong to quickly and accurately respond to instructions and output high-quality answers in complex tasks and scenarios. This is 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 from tens of trillions of data 500 billion word data were cleaned and screened for training large models. Compared with other models, the high-quality Chinese corpus allows "Tiangong" to better understand the Chinese context, vocabulary and grammatical characteristics, more accurately understand the intentions of Chinese users, and is more in line with local users' usage. preferences.

The construction of a large-scale language model has its own technical threshold and is by no means a day's work. This is why there are many comments 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.

The reason why we were able to take the lead in handing over the answer sheet of "Tiangong" is because Kunlun Wanwei's deep cultivation in the field of AI began a few years ago. Kunlun Wanwei began to deploy the AIGC field in 2020, and the birth of the "Tiangong" large 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". "Intelligence Code".

Kunlun Wanwei CEO Fang Han said that Kunlun Wanwei’s business includes browsers, social entertainment, news, games and other sectors, covering more than 70 countries on five continents, and content are very closely related, so they have always been very sensitive to technological progress in content generation. After the birth of GPT-3, the management judged that this was a milestone in the field of content generation, and began to invest in the field of music AI from 2020. Singularity Intelligence realized the future application potential of AI technology as early as 2020, began investing in the field of large models that year, and released a tens of billions of large models in 2021.

In 2022, Kunlun Wanwei will begin to expand from music AI to multi-modal AI. Only by self-developing hundreds of billions of large models can we establish core barriers and seize the initiative. At this time, Singularity Intelligence also became more and more aware that hundreds of billions of large models were a breakthrough for AGI. The two parties hit it off immediately, and it became a natural choice to cooperate and self-develop "Tiangong".

Looking at the future of the large model track, multi-modal pre-trained large models will become a battleground. This is also the only way for the evolution of "Tiangong". The challenge is that image and video understanding consumes more resources and requires more training cards and training resources. Perhaps only players with real strengths in data, algorithms, and computing power can persist until the end.

What are your expectations for the future of "Tiangong"?

The above is the detailed content of Write weekly reports, change codes, and interact for 20 consecutive rounds. The new domestic ChatGPT player 'Tiangong' is here. For more information, please follow other related articles on the PHP Chinese website!