Some people say that the AI model is a game that only a few giants can play.
So far, there are not a few people who agree with this view. Since ChatGPT quickly became popular all over the world last year, it was like a spring breeze overnight, and AI large models were everywhere. Google, Amazon, Ali, Baidu and other giants have joined the "arms race" of AI large models, which seems to indicate that the market is concentrating on large companies.
But is this really the case? Not to mention that foreign generative AI unicorn companies such as Anthropic and Cohere are extremely active in the market, and there are also many domestic AI companies entering the market competition of AI large models from different angles, and the market is showing a scene of a hundred flowers blooming.
Recently, the domestic AI unicorn Yunzhisheng handed over the first answer sheet in the field of AI large-scale models - the large-scale model of mountains and seas. Yunzhisheng has fully demonstrated the ten abilities of Shanhai large model language generation, language understanding, knowledge question and answer, logical reasoning, etc., and Shanhai large model is in the industry's leading test scores in the medical field such as MedQA evaluation and clinical practitioner qualification examination.
Yunzhisheng's practical actions show that the market competition for AI large-scale models has just begun, and the fallacy that only large companies can play well can be put to rest. Just as Huang Wei, the founder and CEO of Unisound, said: "Unsound will continue to upgrade the capabilities of the mountain and sea model. The goal is to match the general capabilities of ChatGPT within this year, and surpass GPT4 in vertical fields such as medical care, IoT, and education."
How to do a good job of AI large model
Throughout the development of the AI model, it has gone through the stage from training the model to training the model. Today, the AI large model scene is regarded by the industry as an important path to realize AGI (Artificial General Intelligence) by virtue of its many advantages such as generalization, generalization, and large-scale replication.
The reason why people are optimistic about the prospects of the giants in AI large-scale models is nothing more than the advantages of the giants with strong financial strength, a large R&D team, and complete infrastructure.
Although these conditions are key factors in the development of large AI models, they are not decisive. In fact, the success of OpenAI just shows that the research and development of AI large models is essentially a combination of multiple algorithms + correct training methods. It tests the comprehensive engineering capabilities of infrastructure + data + algorithms. This comprehensive capability is not a giant exclusive to them.
For example, the training of large AI models is extremely challenging for computing power, but it does not mean that the more computing power piled up, the better, and the stronger the better. As the saying goes, Han Xin counts troops, and the more the better, the AI large model needs a steady stream of computing power, but it needs "Han Xin" who can strategize. Liang Jiaen, founder and CTO of Unisound, said bluntly: "Strong force can produce miracles, but brute force cannot produce miracles."
It is reported that Yunzhisheng is an AI company that paid attention to the construction and investment of computing power infrastructure earlier. Since 2016, it has continuously deployed computing power and other infrastructure, and built a supercomputing cluster Atlas with efficient architecture + full-stack algorithm, becoming the largest in the southeast region. One of the supercomputing clusters. Yunzhisheng’s large-scale model of mountains and seas was launched at the end of last year, and the first version was developed at the end of February this year. It was officially released in just half a year, which is inseparable from the help of the supercomputing cluster Atlas.
Another example is that high-quality data sets and scientific training methods directly affect the effect of large models, which tests the data accumulation, data processing and training data engineering capabilities of AI companies. OpenAI's "Scaling Laws for Neural Language Models" scaling law emphasizes the importance of increasing high-quality data sets for model training. OpenAI CEO Sam Altman also revealed that OpenAI's data set is much larger than outsiders imagined, and a lot of work revolves around data engineering.
Similarly, Yunzhisheng, which has been established for ten years, has rich accumulation in data accumulation and data engineering. Its UniScale training and reasoning integration framework and UniDataOps data optimization parallel processing framework have been tested in various scenarios. Liang Jiaen revealed that in addition to the open source English and Chinese corpus categories, Yunzhisheng also added Chinese and medical data accumulated over the years to train the mountain and sea model.
Therefore, Yunzhisheng was able to complete computing power expansion, algorithm verification, parallel acceleration, data optimization, etc. within half a year, realize the GPT-based architecture upgrade and successfully release the large model of mountains and seas. Huang Wei said bluntly that the AI large model is an "engineering revolution" in the field of AI. It integrates and comprehensively applies various AI technologies in the past, and gradually realizes the large model from quantitative change to qualitative change.
What is the answer sheet handed in by Shan Hai?
As we all know, in the past few months, domestic AI large-scale models have emerged one after another, but their performance is also uneven, and there have even been discussions about various topics such as "Kung Pao Chicken" memes and "Matryoshka ChatGPT".
However, it is normal for the large AI model to answer wrongly, because various factors such as data set quality and training time are involved. Therefore, everyone is very curious, how does the large model of mountains and seas trained by Yunzhisheng in a short period of time perform?
According to Yunzhisheng, the large model of Shanhai has ten capabilities: language generation, language understanding, knowledge question and answer, logical reasoning, code ability, mathematics ability, security compliance ability, seven general capabilities and plug-in expansion, domain enhancement, and enterprise customization Three industry landing capabilities.
Big Data Online also conducted a test on the large model of mountains and seas, including logical reasoning, mathematical operations, coding ability, etc. The overall performance was excellent, and it was able to give clear and accurate answers to logical problems.
Open knowledge question and answer, the large model of mountains and seas can give accurate answers in the case of continuous questioning, and its performance is amazing.
Mathematical operations, the big model of mountains and seas can understand the meaning of the topic very well, and convert it into a calculation formula for calculation and get the correct result.
Logical reasoning, for various logical problems, the mountain and sea model can effectively judge the key to the problem and give a reasonable explanation.
Code ability, Shanhai large model has strong code generation ability.
Human Values Alignment
It is worth noting that the large mountain and sea model also prepares a wealth of plug-in functions such as calculator, weather query, Web search, and discovery of surrounding areas. Users can obtain more accurate information through the rich plug-in functions.
On the whole, the general-purpose capabilities of the large mountain and sea model are quite satisfactory. For example, in terms of open knowledge questions and answers, the Shanhai large model can accurately understand the meaning of the question and answer and give the correct answer, and can answer coherently in the case of continuous questioning; Answers to sexual reasoning questions can be done with ease, and rarely fall into "stuck". A solution that aligns with human values.
Undoubtedly, the outstanding performance of the large model of mountains and seas also gave Yun Zhisheng great confidence. "The goal of the large model of mountains and seas is to match the general ability of ChatGPT within this year." Huang Weiru said.
Open the road to the value of large models in To B
At present, AI large-scale models have been tested in the Internet scene. From ChatGPT becoming the fastest App in history to reach 100 million users, to Internet companies using Internet technology and search engines, Internet companies are generally using AI large-scale models to reshape their own product.
In addition to To C, the expansion of AI large models in the field of To B has also opened the road to the commercial value of artificial intelligence. According to IDC statistics, the penetration rate of China's artificial intelligence in various industries has continued to increase in the past two years. Among them, the penetration rate of telecommunications, finance and manufacturing has increased by more than 4%; the absolute penetration rate of education, energy and medical industries is relatively low, and there is huge room for future improvement. It is estimated that by 2026, the penetration rate of the artificial intelligence industry is expected to reach 20%.
As we all know, from the perspective of the general trend of digital economy and digital China construction, industrial digitization and industrial intelligent upgrading are an important puzzle, among which the application of AI large models in the industry is the most important starting point. Huang Wei said bluntly: The AI big model brings new capabilities to artificial intelligence, and can create more products to meet the needs of industry users for intelligence.
In fact, AGI has broad application prospects in industry scenarios, but it is by no means easy. In the author's opinion, Yunzhisheng's U+X strategy (U: AI technology and product capabilities, X: industry application scenarios) has opened up a value path for the deep cultivation of AI large models and AGI in the industry.
The author believes that the advantage of Yunzhisheng's U+X strategy is that it can connect data, applications and models to form a flywheel. First find a suitable scene in the industry, then accumulate specific effective data through user behavior feedback, and then feed back the AI large model; the continuous iteration and optimization of the AI large model is conducive to the further application of AGI in the industry and the collection of more Effective data, thus forming a flywheel effect, the effect of the larger model will be better later.
Taking the medical field as an example, it is an industry with high complexity and knowledge density. Yunzhisheng's U+X strategy has been deeply cultivated in the medical industry, which makes its mountain and sea model really amazing in the medical field. At the press conference, UniSound demonstrated multiple scenarios such as assisting doctors to generate complete medical record plans based on medical records and assisting medical related claims with the assistance of the mountain and sea model, which greatly improved business efficiency and intelligence. In addition, in the MedQA evaluation, the Shanhai large model ranked first with 81.56%, surpassing GPT4, Med-PalM and other models; in the clinical physician qualification examination, the Shanhai large model scored as high as 511 points, far exceeding the average score and passing line .
It is reported that the large-scale model of mountains and seas will be successively launched in top tertiary hospitals in China. It is conceivable that with the application of the mountain-sea model in more medical scenarios, and by learning more medical data, it is expected to further enhance the application scope and depth of AGI in medical scenarios in the future. In addition to the medical field, Yunzhisheng also demonstrated the application of the large mountain and sea model in sales, knowledge management, education, smart IoT and other scenarios.
"In the long run, AI large models will become the 'power plants' of the AI 2.0 era. Big data is the fuel, computing power is the boiler, and algorithms are the generators. Using large models in the future is as convenient and convenient as using electricity." Huang Wei added.
Why are you optimistic about Yunzhisheng
At present, big AI models are flocking to each other, including giants such as the Internet and cloud computing, as well as newly established AI companies, and even unicorns like Yunzhisheng who have been deeply involved in the AI field for many years. It is foreseeable that the competition in the AI large-scale model market will accumulate more and more in the future.
So, in the face of fierce market competition in the future, what is the prospect of Unisound? In the author's opinion, judging from the prospect of the artificial intelligence market, Unisound's strategy and core capabilities, Unisound is expected to make continuous breakthroughs in the market and achieve self-transcendence.
First of all, under the background of the digital economy and the construction of digital China, it provides Yunzhisheng with a very broad market stage. IDC predicts that the scale of China's artificial intelligence market will exceed US$14.7 billion in 2023, and will exceed US$26.3 billion by 2026. The market growth will mainly come from the replacement of AI applications built in the past few years and generative AI by applications based on large models. The incremental market and brand-new AI-powered enterprise-level applications brought by
Secondly, the U+X strategy that Unisound has adhered to for the past ten years has proved to be a difficult but correct road to AI commercialization. With the advent of the AI large-scale model era, Yunzhisheng's U+X strategy will undoubtedly be even more powerful, which will help accelerate the commercialization of AGI in the industry, and is expected to open up a broader industry market.
Thirdly, the two amazing items of the large mountain and sea model prove that Unisound's deep data accumulation and engineering capabilities are first-class in the industry, which is also the core advantage of Unisound's ten-year practice accumulation. As OpenAI embarks on a road of AI large-scale models, Yunzhisheng is expected to keep up with or even surpass it by virtue of its core advantages such as data accumulation and engineering capabilities.
Comprehensive observation shows that the AI big model is entering the Warring States period, and the market competition pattern is far from determined. The AI unicorn Yunzhisheng created a large model of mountains and seas in just half a year, which fully proves that Chinese AI companies can also fight. Facing the future, AI large-scale models and AGI have a tortuous but bright road to industrial application, and Unisound is expected to rely on its own advantages to enable large-scale mountain and sea models to break through the clouds and set sail for long-distance sailing.