Big Models Need Doers

Original: Wu Xianzhi Wen Yehao

Editor: Wang Pan

Source: Photon Planet

Image source: Generated by Unbounded AI‌

The domestic large-scale model war was once shrouded in many doubts. With JD.com and Tencent taking their seats one after another, this complicated game is becoming clearer.

Baidu and Ali, as the first batch of large-scale model players in China, won the early traffic while “getting the head start”. In contrast, although Tencent and JD.com started a little later, they are more able to examine the large-scale model track in depth to deepen their understanding of the industry.

Because of this, some major manufacturers that released large-scale models relatively late did not blindly follow the public and invested in the tide of general-purpose large-scale models. For example, Huawei’s large model mentioned practicality before, and JD.com has chosen a more farsighted and more targeted path by virtue of its deep accumulation in the industry and supply chain. At the same time as the JD Cloud Summit (JDD Conference), an industry-oriented JD Yanxi model was launched.

In the current era of general-purpose large-scale models, the Yanxi large-scale model and its industrial value creation path launched by JD.com are not only an excellent attempt to establish a large-scale model business order, but also bring a new thinking and direction to the entire industry. In the stagnant waters of commercialization of large models, industrial large models that are closer to landing are likely to lead the way in the future competition for large models and become a new force that cannot be ignored.

Large model, neither a “story” nor a “toy”

Every wave of technological succession often carries a lot of sediment.

From chips, robots, to AI in the early years, in every wave, there are always some players with ulterior motives mixed in, mixing technology and marketing together, making the originally clear track even more muddy. The same is true for the domestic large-scale model tracks at this stage.

On one side of the track, players are struggling to find out the technical context of the large model, and trying to find the direction of landing; on the other side, even companies that have nothing to do with the technology field are leaving the field one after another, training out The so-called “big model” products - I have to admit that the current situation of “everything can be a big model” has the meaning of “everything can be a metaverse” two years ago.

Obviously, when the so-called “big model” becomes a synonym, serving its own “storytelling” appeal instead of creating actual value, then there is a high probability that it can only go to the fate of the metaverse. And this is also true for those players who honestly study technology. After all, technology itself is difficult to directly create value - the end of the large model is not to train the general large model itself, but to let the technology generate value, and then realize mature commercialization.

Therefore, the seemingly funny operations of the chasing fans actually sounded a wake-up call for the large model track. After all, the ChatGPT that ignited the AIGC has subtly affected the players’ view of the large model, causing many players to rush to the general large model and launch a variety of “variants” of ChatGPT.

Objectively speaking, the general-purpose large model has its value, but at the moment when the competition dimension is becoming more and more fierce, the general-purpose large-scale model is not smooth:

On the one hand, players flock to a single field one after another, and are very likely to fall into the situation of “repetitive manufacturing of wheels”. It is conceivable that it is difficult to break out of the battlefield full of domestic and foreign technology giants.

On the other hand, a general-purpose large model is a typical product of the separation of consumption and payment, and it will become a “toy” if you are not careful.

To give a simple example, a wide range of C-end users may ask questions or even discuss the universe and the sky, but the vast majority of ordinary users do not actually have productivity demands. After a short-term intensive experience, they will quickly feel the freshness of emerging technologies. Lost, may not have the power to use for a long time.

Based on this, even if the current general-purpose large-scale model can improve the efficiency of content creation to a certain extent, except for the cost reduction and efficiency increase within some content industries and organizations, it has not yet developed a mature and replicable business model. It is foreseeable that as the general large-scale model track becomes more and more crowded, players are bound to face many challenges in terms of commercial extension.

In the final analysis, now that the “singularity” has come, large models are not only a milestone in the succession of technology, but also a key driving force for shaping the future. According to this logic, the large-scale model war at this stage is by no means a short-distance race, but a systematic project. If players want to go through the cycle and get to the end, they can’t just rely on a single point of breakthrough at the technical level, but need to simultaneously think about many dimensions such as technical direction, scene application, and business model.

Industrial large-scale models, a new front in the “Hundred Models War”

From 1997, when “Deep Blue” defeated the chess master Garry Kasparov, to “AlphaGo” entered the Go circle, and then to visual system AI and automatic driving, AI has experienced many rounds of exciting evolution in the past. The wheel seems to be standing on the verge of application explosion, but the flower buds that are full of branches have not been able to bloom for a long time.

The main reason behind this is that the technology has not yet formed a deep application in the industry. After all, the end of technological progress is not trapped in the laboratory, but a dive into the “real world”.

Following this logic, looking at today’s general-purpose large-scale model tracks, there is still a long way to go before taking root in real business scenarios and creating real value.

Xu Ran, CEO of Jingdong Group, said at the JDD conference that the large model itself is a tool to realize industrial value, not an end. The real value of large models must be realized in industrial applications.

In other words, the big model is not the goal, but the application is the goal.

The current large model manufacturers often regard model parameters as the test standard for the quality of large models. As everyone knows, at the level of commercial implementation, huge parameters also correspond to high costs, and there are also problems such as long corresponding time and poor concurrency.

A simple example, some “parameter monsters” cost two or three cents to answer a question, and have to wait for 5-10 seconds. No matter how accurate the answer is, it is difficult to achieve large-scale commercialization. In addition, the current general-purpose large model has an accuracy rate of about 85%, which may be sufficient for ordinary users, but in serious business scenarios, this error is likely to have an impact that cannot be ignored on the business.

Regarding the application problem, the technical leaders of several business lines of JD.com mentioned that people would laugh at a fabricated answer from GPT, but once it is implemented in the actual application process, any deviation will lead to huge losses.

Dr. He Xiaodong, Dean of JD Exploration Research Institute and President of JD Technology’s Intelligent Service and Product Department, once experienced something personally, which is quite representative. “A large model answers the square root of 143, and the answer given is 11.5 (actually approximately equal to 11.96). If it is used in an actual application scenario, this answer will bring huge losses.”

In the technical field, model parameters and accuracy are important, but in the business world, the key is that the large model itself is easy to use and stable. In this regard, the industrial model closely related to subdivided industries undoubtedly has natural advantages.

However, it is not easy to develop a large-scale industrial model. As we all know, training data is the basis of large model learning, and also determines the generalization ability and application scenarios of large models. Therefore, in addition to breakthroughs at the technical level, first-hand scenarios and data from the industry are equally important for the development of large industrial models.

Taking JD.com as an example, the reason why it launched a large-scale model for the industry is largely due to its strong industrial genes. After all, among the large domestic manufacturers, JD.com, which links the consumer market and the two ends of the supply chain, has a strong connection with the same industry and possesses a large amount of high-quality data.

It is reported that when the Yanxi large model is trained, 70% of the general data and 30% of the original data of the digital intelligence supply chain are integrated. It can be seen that JD.com does not purely emphasize parameters, nor does it deliberately tell “story”, but focuses on the “tuning” level, aiming to create a large model that is highly integrated with the industry.

The large-scale industrial model may become an important step towards large-scale commercialization of the large-scale model track, and players who gradually understand the logic are gradually getting on the car.

Recently, Tencent, which has been holding back for a long time, has released its own large-scale industry model; Baidu, which holds high the banner of general-purpose large-scale models, has also released large-scale industry models covering transportation, energy and other fields. It is not difficult to see that as the giants have increased their size, the large-scale industrial models that are closer to commercialization have become a new front in the “Hundred Models War”.

Meet on a narrow road, the one who gets the “scene” wins

Whether it is a general-purpose model or an industrial model, the construction of a new business order cannot avoid “scenes”.

In other words, for implementation, obscure technical terms and dazzling commercial PPTs are all castles in the air. Only by truly applying the ability of large models to the scene and generating actual value can a virtuous circle be opened.

At the moment when the demand for AIGC in all walks of life is blowing out, it is not difficult to find the so-called application scenarios for large models. But if you want to find a scene suitable for large-scale landing of large models and run through the commercialization path, you may have to go through a lot of detours.

Following this logic, the players who have already presented the big model are constantly trying in various subdivisions, trying to find their own foothold:

Baidu seized the hot spots of the annual college entrance examination and launched AI volunteer assistants. While making full use of the capabilities of large models, it also tried to use this to enter the C-end market; Ali used Tmall Genie as an anchor to explore the application of large models in the field of consumer electronics. Expansion capacity.

JD.com, which develops large industrial models, puts forward the formula of “value of large models = algorithm × computing power × data × square of industrial thickness”, and the so-called “industrial thickness” is just piled up by specific scenes. become.

Regardless of the path, large-scale model players are bound to continue to try and make mistakes before reaching the “Promised Land”, and even cross the “Red Sea”.

Faced with the difficulty of landing large models, some players choose to play the role of “water sellers” to help companies build their own large models. However, JD.com, which has always been pragmatic in its style of play, has proposed a “three-step” strategy, that is, first build a general-purpose large model, then explore scenarios and applications internally, and then gradually open up its capabilities to the outside world-while using itself as a test field, it is also self-developed. Digest the cost of trial and error to ensure that large-scale model products can create real value.

It is reported that within JD.com, the large model has not only been embedded in common application scenarios such as digital marketing, operation process optimization, and customer service, but has also been extended to many vertical scenarios such as retail, logistics, finance, and health.

Taking the logistics field as an example, in the face of this complicated system engineering, JD.com has explored multiple exploration paths: the supply chain product Jinghui, which has been built for 5 years, in addition to rich native algorithms such as AI prediction and operation optimization, through Open ecological technology can not only communicate well with the algorithms and data of heterogeneous systems, but also be more expressive in terms of sales forecast, inventory, supply and replenishment planning with the support of large models. At the same time, due to the wide application of AIGC , its interactive supply chain control tower can help users quickly locate and solve supply chain problems.

In the fund management scenario, JD Finance’s “smart base selection” product was launched. Traditional fund screening is costly to understand and cumbersome to operate, which directly affects the success rate of transactions. With the help of large models, JD.com has optimized the links of intent matching, algorithm generation, intent recognition, and multiple rounds of dialogue, making the accuracy rate of common screening questions reach 90%, effectively improving customer experience and transaction efficiency. This product will also fully serve financial institutions in the future.

It can be seen that JD.com’s “three-step” strategy has achieved initial results, and has gradually penetrated into the texture of vertical industries such as logistics and finance. It is foreseeable that as the strategy is gradually rolled out, JD.com will also continue to accumulate landing scenarios and high-quality data, thereby turning the flywheel for the commercial landing of large-scale industrial models.

In the final analysis, the seemingly complicated large-scale model warfare must return to a core issue, that is, how technology can bring actual value to the real industry. At this stage, players with different paths can only gradually explore and practice the answers to the questions in this long marathon.