At the press conference of the large model of mountains and seas, Huang Wei had an impressive sentence, "In the past ten years, Yunzhisheng was born for mountains and seas.
Author | Pi Ye
Produced | Industrialist
"Who can make the best large-scale model in China?" At an internal sharing meeting of an investor in Beijing in March this year, someone raised such a question with great anticipation. But unfortunately, there is no answer to this question, or no one can prove their own answer.
In the past few months, with the emergence of ChatGPT, a phenomenon-level product, the popularity of large models has only increased. Whether it is a major Internet company, an emerging Internet entrepreneur, or a traditional AI company, the Large models have sprung up in the market.
According to incomplete statistics, since March this year, more than 30 large-scale model products have appeared in the Chinese market. In addition to the inherent general-purpose large-scale models, there are also medium-sized models (industrial models) specifically for subdivided tracks, such as low-code , industry, energy and more.
But two months have passed, and the question at the beginning of the article remains unanswered.
The most realistic questions are that with the increase in the number of large models, the market has a more realistic or prudent attitude towards large model products, that is, what problems can large models solve? In addition to intelligent dialogue and logical reasoning, how far is the distance from the large model to the real industrial scene? How long does it take to transform into real social productivity?
Behind the question is not only the deep thinking of the market on large-scale model products, but also the exploration of the current direction of China's digital and intelligent future. The companies behind the large-scale model products need to use more practical actions to complete self-certification.
Recently, industry experts connected with Dr. Liang Jiaen, the founder and CTO of Unisound, to understand what this wave of large models means from his perspective, and for Unisound, a veteran of the AI track, it What kind of answer sheet is being handed over.
According to him and Yun Zhisheng, change and self-proof are already on the way.
1. Yun Zhisheng, "Hand in the test paper"
511 points—this is a score announced by Huang Wei, the founder of Unisound, at the release site of Unisound's large model. It is the score that Yunzhisheng Mountain and Sea Large Model can get in the medical examination for clinical practitioners. The total score of this test is 600 points, and the average score of candidates who take the test is 365 points.
In the medical field, the MedQA evaluation also exceeds 81 points, which is a report card exceeding GPT-4.
"We want to train the large model from undergraduate training to doctorate in the professional field through domain enhancement training." Liang Jiaen told the industry. In the medical field, this idea is being implemented.
This is exactly the difference in this Yunzhisheng mountain and sea large-scale model conference. That is to say, in addition to the language dialogue ability and logical reasoning ability displayed on-site by the large-scale models on the market, the voices of the large-scale models of Shanhai are also concentrated on the industrial level, such as medical care, home furnishing, education, automobiles and other industrial fields.
Taking the medical field as an example, people can not only obtain accurate medical answers based on a certain disease, but in clinical practice, doctors can also generate medical record keywords through voice, and with the assistance of the mountain and sea model, they can assist doctors to generate a complete medical record plan based on the medical records ; At the same time, it can also help patients and insurance companies to make medical-related insurance claims.
Another example is in the field of the Internet of Things, the mountain and sea model can realize active intelligence, emotional responses, and multiple rounds of complex dialogues, etc., helping people arrange their schedules like "Jarvis" in Iron Man and becoming a smart property steward.
In addition, the Shanhai large model can also become a "sales expert", "knowledge management expert", "speaking expert" in a specific field, etc., and deeply empower specific industrial scenarios based on the large model.
"Basically, we have already trained the open source English and Chinese corpus categories listed by organizations such as OpenAI, and we have also added our own Chinese and medical data." Liang Jiaen told us .
According to Yunzhisheng's "U+X" strategy, the answer sheet handed over by Shanhai large model is not only on the "muscle" of the general large model, such as conventional language generation, language understanding, logical reasoning, data and code capabilities, security Compliance capabilities are also more specific to the implementation of industries, that is, through plug-in expansion, domain enhancement, and enterprise customization, more targeted implementation and adaptation can be achieved in professional industrial fields.
What kind of answer sheet is this?
2. From dedicated to general purpose, AI to B behind the path
"For example, if OpenAI can achieve 95 points in general, but the reliability of many professional fields cannot meet the practical requirements; our general-purpose base can achieve 90 points, and then we will give priority to strengthening in different fields, and finally can be in the field. Landed in the ground." Liang Jiaen said.
With the continuous fire of OpenAI, while people marvel at the turning point of AI, some hidden problems are also emerging. For example, for the GPT model, its current value is more in the general sector, such as the understanding of semantics , logical reasoning and other abilities, but on the specific industry side, it must undergo targeted training to meet the basic requirements.
Among them, medical treatment is the field that people talk about the most at present. As a "serious" discipline and field, its medical treatment has zero tolerance for the "illusion" phenomenon that appears during the training process of the large model. To land in the field, sufficient professional data training and fine-tuning must be carried out in order to achieve "industrialization".
It can also be said that if the financial industry is a highland where the capabilities of domestic manufacturers such as databases are self-proven, then medical care, as an industry with extremely high complexity and knowledge density, can be regarded as the strongest whetstone for large models to have the ability to empower industries .
"Prior to this, we have done a large number of medical intelligentization cases, with a large amount of medical data accumulation and comprehensive knowledge mapping capabilities, which are the basis for us to pre-select the medical direction." Liang Jiaen told us.
This is the direction Yun Zhisheng has always chosen to attack. Take Yunzhisheng's "medical voice input system" as an example. It allows doctors to input text into the place they want to input in real time by speaking, solving the inherent problem of "recording medical records while communicating" and freeing doctors from the burden of transcribing medical records. time.
It is understood that the voice recognition accuracy rate of this system is over 95%, especially in neurology, immunology, hematology, general internal medicine and other departments with many patients with intractable diseases, and the voice recognition rate of individual departments even exceeds 98%. %.
Similar cases of medical intelligence have been one of Yunzhisheng's main directions in the past few years, and these intelligent solutions and data are now trained into the mountain-sea model. It is understood that in the future, the large-scale model of mountains and seas will be successively launched in the top three top hospitals in China.
Similar to medical care, the Internet of Things is also one of the key landing directions of Yunzhisheng's large-scale model of mountains and seas. Based on the ability of the large model, Yunzhisheng can achieve Smart IoT 3.0 on the basis of the original AIOT ecosystem. Similarly, this capability has now been gradually implemented in some scenarios such as smart industrial parks.
The direction corresponding to medical care and the Internet of Things is exactly the path of large-scale models that Unisound is taking—from dedicated to general. That is to say, compared with continuous general-purpose data superposition on the base of a general-purpose large model, Yunzhisheng’s approach is to conduct targeted training directly in special (industry) fields on the basis of the underlying general-purpose base, “gradually integrate each field They are all at the level of a doctor's degree", and finally feed back the large model of the base to achieve a more accurate industrial expression.
In fact, this is exactly the path of large-scale models that is most in line with the landing of the current industry. That is to start from the problems of the industry and the actual situation, carry out the corresponding industrial model expression, through the accumulation of model capabilities in each professional field, and finally achieve the "generalization" and "professional reliability" of the large base model.
In addition, with the continuous influx of industrial data, Yunzhisheng uses the optimized framework and combed high-quality data to ensure that the large model has the attribute of "anti-bloat", while realizing the accuracy of the model and the ability to serve the industry.
"The more data is not the better, and everyone in the academic circles is also discussing that after a certain amount of data is reached, the accuracy and effect of the model will not be continuously optimized as the amount of data increases. Data quality and diversity are more critical .” Liang Jiaen told us, “In the end, we still have to return to the optimization of the framework and data.”
If you turn the timeline forward, as early as 2016, the Unisound team began to build a large-scale supercomputing platform Atlas. On this platform that represents the Titans in Greek mythology, Unisound began AI empowerment attempts in industries such as medical care and the Internet of Things.
"Therefore, it is not difficult for Yunzhisheng to simply pile up data now. The difficulty lies in making it reliable and usable in each specific industry." Liang Jiaen said.
3. In the era of large models , the released "AI power"
If you look at the path chosen by Yunzhisheng from a larger perspective, you will find that everything is not accidental. Its outstanding features are its emphasis on engineering optimization and industrial scale capabilities, and it is not an easy path.
For example, when the underlying Atlas platform was established in 2016, Unisound was only a four-year-old startup company. "Experts in Silicon Valley were even surprised that it was too early for a four-year-old startup company to consider this issue."
For example, the Data Center Model Optimization (DCML) layer on top of the Atlas platform. After Yunzhisheng started to enter the medical industry in 2016, it gradually realized that solving the data differences of different hospitals and departments is the key to realizing large-scale applications, and began to build a development model of "unified model architecture + data iterative optimization". That is to say, AI capabilities can be applied to various fields in a more standardized manner, and then the pan-"standardization" of AI capabilities can be realized. This is the DCML platform that supports the training of large mountain and sea models.
For another example, also in 2016, Unisound optimized the deep learning technology, which is generally believed in the industry to be able to run on the GPU, to run on the CPU or even the mobile phone chip. At the same time, they even made the model into a WiFi chip with a main frequency of only 200MHz and a computing memory of only 200k.
These layouts, which seemed "unreasonable", "advanced and laborious" at the time, now constitute Yun Zhisheng's confidence to hand over its answer sheet in this new trend of large-scale models. Whether it is medical care, or home furnishing, or now that Yunzhisheng is going deep into the "education" and other industries, the Shanhai large model can be quickly adapted and implemented based on complete technical support.
Liang Jiaen told us that now that the large mountain and sea model is deployed locally, enterprises only need to use "A10"-level GPUs to achieve the inference effect of "A100/A800". Tens of thousands of dollars, the cost of local deployment can even be reduced by more than 80%, which can be further optimized in the future.
In addition to technical support, based on past experience in serving medical and other industries, Yunzhisheng can build knowledge maps in other fields more quickly, and cooperate with large models to achieve more professional implementation. This is the path that all AI companies are practicing, but there are not many teams with both large model and knowledge graph capabilities.
In Liang Jiaen's words, "Now Yunzhisheng is undergoing the third technological upgrade. " If we say that in 2016, Yunzhisheng completed the full stack from "sound (perception)" to "knowledge (cognition)". The technical system is upgraded, so starting in 2022, what this AI company is completing is the upgrade from "dedicated AI to general AGI".
However, today is different from what Yunzhisheng faced in 2016.
That is to say, in 2016, Yunzhisheng was making innovative attempts no matter from the underlying supercomputing platform, the upper-level data model, or the deepening of the medical industry, and verified the company's own path forward. But at that time, whether it was the functions that AI could achieve, the reusability of different industries, or people's cognition of AI, they all remained at a superficial level.
It can be understood that the production value of AI in the past was more of an assembly model of "dedicated business system + standard AI parts", the standardization and reusability of the business layer were quite insufficient, and there were bottlenecks in understanding ability and flexibility; but now With the emergence of large models, most of the tasks can be connected through natural language, the IQ has been significantly improved, and business capabilities can be expanded through large model tuning instead of programming. The MaaS model has become possible. It is redefining AI in the digital transformation of enterprises The position and role in the book are more specific and practical. For AI companies, it also corresponds to more powerful and reusable product technology value.
In other words, the large model breaks the market's cognitive barriers to the business boundaries and commercial value set by AI companies such as Yunzhisheng, and re-recognizes the large model that integrates cognitive capabilities such as language, knowledge, reasoning, and decision-making. The core value and significance of intelligence, what they can do, and the imagination space that can be achieved in the future are all being updated.
And these foundations are all based on the technical persistence and market exploration of companies like Yunzhisheng in the past many years. At this moment when AI is becoming the core productivity of society, Unisounds are becoming the real protagonists.
At the conference site of the large-scale model of mountains and seas, Huang Wei had a very impressive sentence, "In the past ten years, Yunzhisheng was born for mountains and seas."
Write at the end:
Yunzhisheng's large-scale model path can be summarized as "industrial AI". That is to say, the path of many large models on the market is to first carry out the underlying training of large models, and after training to a certain order of magnitude, look for industrial scenarios and values that can be implemented.
However, Yunzhisheng's purpose and direction are very clear, that is, the initial direction is to enhance and implement a specific (industry) model capability on the basis of a general-purpose large model, and first verify it in advantageous industries such as medical care and the Internet of Things. Then expand different application scenarios and try training separately. The criterion is whether the model is usable, reliable, and valuable in the industry.
Judging from the current stage where the large model has not reached the end, although it is impossible to discuss whether the path is right or wrong, the paths of companies such as Yunzhisheng can make the large model intersect and connect with the real world, and thus transform step by step into The real and visible productivity not only completes the presentation of TOC, but also realizes the "non-bubble" demonstration on the industry side. On the specific medical industry side, its effect and capability currently exceed GPT-4.
This is what Yunzhisheng brought us to think, and it is also a new self-proof of China's AI technology.