"Will Source 1.0 replace journalists?"
This is indeed a tricky issue. "Source 1.0" is the world's largest Chinese AI huge model, which can poetry, talk, write novels, write news, and arouse the mixed feelings of dozens of reporters on the scene. But Liu Jun did not directly answer this question: "Li Baidou wine hundred poems, Jia Dao two sentences three years." In the future, 'Source 1.0' will become everyone's 'assistant' and will become an 'assistant' like Li Bai and Jia Dao. ”
Indeed it is.
As Liu Jun, vice president of Inspur Information and general manager of inspur information ai&hpc product line, said, artificial intelligence is an "assistant" and "tool", and Inspur hopes to think on the shoulders of artificial intelligence, and also hopes that the technology industry can stand on the shoulders of artificial intelligence to complete innovation. During aicc 2021, inspur Artificial Intelligence Research Institute officially released "Source 1.0", which is currently the world's largest Chinese AI huge model, with a parameter scale of up to 245.7 billion and a Chinese dataset of 50,000gb for training.
Looking back now, inspur released a "huge model", which should have been laid out earlier. In 2018, The Wave has made the assertion that computing power is productivity. In 2020, Inspur has successively proposed concepts such as "intelligent computing center" and "meta-brain ecology", which has landed a "huge model" and laid a foundation for computing power and ecology.
This is followed by "Source 1.0".
Although the foundation has been laid, there is a gap between computing power and productivity. This is the gap between perceptual intelligence and cognitive intelligence, from specialized intelligence to general intelligence. Because of the current artificial intelligence, it is still difficult to understand the relationship between the scene and the object, and it is also difficult to form a loop interaction between the microscopic and macroscopic.
Enterprise users and ecological partners are even more unable to do this. They pay attention to applications, but they cannot afford the cost of tens of thousands of GPUs, nor can they shield the complexity of multiple computing power; they also pay attention to algorithms, but they are also unable to complete the collection and training of t-level data.
The emergence of the "giant mode" provides a direction to solve the problem. In 2020, OpenAi launched the GPT-3 model, with a parameter scale of 175 billion, which will push language models, and even artificial intelligence, to a new height. Since then, Google, Microsoft and Zhiyuan and other institutions and enterprises have also launched "big models", and NVIDIA and Microsoft have launched mt-nlg with model parameters of 530 billion.
The same goes for Wave "Source 1.0". It is a complex giant system and an algorithmic infrastructure necessary for enterprise users and ecological partners. "Source 1.0" will promote artificial intelligence and achieve key breakthroughs in small sample learning, generalization ability, constrained reasoning, logical reasoning, etc.
For example, traditional artificial intelligence is limited to a turnip and a pit, and a data model carved out can only be landed in one application scenario, but the enterprise application scenario is not only hundreds or thousands. The small sample learning supported by "big mode" solves this problem, and the most versatile model is trained with the least amount of data and is suitable for different scenarios.
Even so, "huge" is not the pursuit of the wave. Silicon Valley entrepreneurial circles have the concept of "ten times thinking", and new products are ten times better than others. Perhaps, this leads to the misunderstanding - the "tenfold thinking" represented by big models, large computing power, and big data is promoting the AI industry with "violent aesthetics".
But this is not the core logic of the development of the AI industry. Fragmented "small models" are powerless to promote artificial intelligence. The "small model" cannot show good value in terms of small sample learning and generalization ability, and only with the gradual growth of the parameter scale, the relevant capabilities will be continuously improved.
Of course, the "aesthetics of violence" is indeed an aesthetic. If you compare with ten years ago, the starting memory is only 16gb, the main frequency is only 1ghz of the iPhone 4, the current iPhone 13 is indeed "violent aesthetics", but think about the development of the industry ten years later, perhaps the current "violence" will also be regarded as too gentle, perhaps then will sigh: it is the computing power that limits our imagination, the model limits our imagination.
Not only that
In the "Turing Test" of "Source 1.0", the dialogues, novel continuations, news, poems, and couplets generated by the "Source 1.0" model are mixed with similar works created by humans and distinguished by the crowd. The test results show that the success rate of the crowd to accurately distinguish between people and "source 1.0" works has been less than 50%.
This is the power of "Source 1.0".
But "Source 1.0" not only has poetry, but also the far side of the industry. "Source 1.0" focuses on the field of natural language processing, involving typical application scenarios such as: automatic report generation, intelligent assistant, intelligent customer service, intelligent operation and maintenance, text recognition, text search, intelligent translation, etc.
Look at the hotel lobby, only sell cute "vase" robots; listen to the intelligent customer service, full of "routine" questions and answers. You know how much "moisture" there is in natural language understanding today. Explained in professional language, sitting across from you is not necessarily a "dog", but it must be a "computer system that boots in a logical order", it is just following the predetermined steps, performing "logical boot".
Change will also come from this. "Source 1.0" will speed up key technologies such as in-depth question-and-answer, multi-round dialogue, and memory network, and the speed of tempering and polishing will also promote the replication speed of scenes. After one training, the model can be used to perform different tasks, just like the model that learns to write poetry, it will also be able to write novels, models that can write prose, and will also write papers, which are all driving the development from specialized artificial intelligence to general artificial intelligence.
meanwhile.
The open source open plan based on "Source 1.0" was also launched at the same time. Through open source and open AI algorithm infrastructure, Inspur will further enhance the AI research and innovation strength in the field of natural language understanding, so that more industries can obtain the production mode and production efficiency changes brought about by the huge model technology.
And this time, the wave opened up quite thoroughly. At the beginning of the plan, Inspur will provide comprehensive openness from data sets to APIs and source codes for universities and scientific research institutions, meta-brain ecological partners, and intelligent computing center partners. This is obviously a complete AI ecology that covers "production-learning-research-use".
"Source 1.0" has been almost 5 years, and Chinese read all the content in the Internet. Through the self-developed text classification model, it obtained a 5tb high-quality Chinese dataset. At the same time, "Source 1.0" also read about 200 billion words, which is equivalent to the amount of reading per person in 10,000 years. Even "bookworms" like Zhang Taiyan and Wang Guowei could never reach such a height.
Now Inspur has opened up all these accumulations, and the algorithm infrastructure has been completely opened to the "production-learning-research-use" ecological partners. That is to say, supplemented by such a computing power infrastructure as the Intelligent Computing Center, the technological gap, the talent gap and the ecological gap between computing power and productivity will be filled, and the cooperation between "production-learning-research-use" will eventually be coordinated with "source 1.0". At this time, the wave will also truly become the "source" of AI.