news

since chatgpt completely ignited the opening show of artificial intelligence (ai) last year, people all over the world have discovered that no technology has ever been as fast-paced as ai: both large models and chips are being upgraded and iterated at an astonishing speed; and china and the united states have led the entire industry to present a grand scene of thousands of ships competing. with the rapid iteration of technology and the profound changes in the market, the opening show will eventually come to an end, and the real problem will eventually be put in front of everyone: how to apply it? how to make a profit?

as a result, a clear consensus is gradually forming: the development of ai has gone beyond the simple "big model arms race" stage and has officially entered the second half where "deep mining of scenario applications" and "leading innovative concepts" are equally important.

in this new stage of development, how to find the best balance between technological innovation and commercial application? how to seek cooperation opportunities in fierce competition? how to discover and seize development opportunities in challenges? these questions have become key topics that practitioners have carefully considered and answered. as leaders, china and the united states have undoubtedly attracted the attention of the world. the strategic choices, technological investment and market dynamics of the two countries are shaping the development trajectory and competitive landscape of global ai in an unprecedented way.

1

us perspective: investment fog and commercialization dilemma

the united states, as the birthplace of ai technology, has invested heavily in the research and development of large models, but its commercialization process has faltered and faces the risks of low return on investment and market bubbles.

under the ai ​​boom, the market showed great enthusiasm. nvidia, meta, tesla, amazon, google's parent company alphabet, microsoft and apple, the seven technology giants known as the "magna 7", shined in the stock market in 2023, with their stock prices soaring 239%, 194%, 102%, 81%, 59%, 57% and 48% respectively.

however, behind this craze, there is a shadow that cannot be ignored. at the end of june this year, an article by goldman sachs titled "too much investment, too little return" brought the ai ​​bubble theory to the forefront. the article bluntly stated that large companies plan to invest huge amounts of money in ai-related fields in the next few years, but apart from slightly improving the work efficiency of developers, no other significant results have been seen. sequoia capital also believes that the bubble in the ai ​​industry is intensifying, and the demand gap for ai profitability is expanding. many institutions have issued warnings: the research and development of large ai models is facing unprecedented challenges. although capital investment continues to increase, the commercialization process is far from expectations, revenue growth is weak, and profit prospects are worrying. more seriously, the risk of industry bubbles is becoming increasingly prominent, and the low return on investment has caused deep concerns about the sustainability of the ai ​​industry.

specifically, the first problem to be addressed is commercialization and input-output. at present, investment in the research and development of large models continues to increase, but the commercialization path has not yet been clarified, resulting in a serious imbalance in the input-output ratio. the development prospects of the track are unclear, which makes both investors and developers face huge uncertainties. in addition, the insufficient revenue growth and poor profit prospects of ai large models have further exacerbated the anxiety in the market. in particular, with the advancement of technology, the training cost of ai models continues to rise. for example, the "2024 artificial intelligence index report" stated that the training cost of openai's gpt-4 model is expected to be as high as us$78 million, and google's geminiultra model is as high as us$191 million. such high costs are a major burden for many companies, especially when the returns are unclear.

on the other hand, the actual development of ai applications is also worrying. although large models have made significant technical progress, their application has been much slower than market expectations. although ai technology has surpassed human performance in tasks such as image classification, visual reasoning, and english comprehension, it still lags behind humans in more complex tasks such as competitive mathematics, visual common sense reasoning, and planning. this lag in development has not only affected market confidence, but also exacerbated anxiety within the industry, causing people to re-examine the development path of ai technology.

against this backdrop, the us ai field is undergoing a profound reflection and transformation, from a pure technology competition to an in-depth exploration of commercialization paths, and from blind optimism about large-scale model development to an urgent need for the implementation of scenario applications.

2

the increasingly expensive track urgently needs to seek new directions

in the surging tide of artificial intelligence, big models have undoubtedly become dazzling stars. their powerful generation capabilities and broad application prospects make them a hot investment project. however, the current situation of the big model industry is like "ice and fire". on the one hand, there are rapid technological breakthroughs and surging investment booms, and on the other hand, there are many challenges and difficulties faced by enterprises in practical applications.

since chatgpt came out, large model technology has developed at an astonishing speed, constantly broadening the cognitive boundaries of artificial intelligence. technology giants such as google, microsoft, and meta have increased their investment in the field of ai and competed to build data centers. the battle for computing resources has become increasingly fierce and has become the focus of industry attention. however, the report of the wall street journal on march 31 was like a cold water pouring on the head, revealing a cruel reality: the ai ​​industry spent as much as $50 billion on nvidia chips alone to train large models last year, while revenue was only $3 billion, showing a huge gap between high investment and meager returns.

the scale of ai spending by technology giants also proves this point. according to media reports, meta expects ai spending to surge to $10 billion this year, google invests about $12 billion per quarter, and microsoft spends $14 billion in one quarter, indicating that this spending will continue to grow significantly. with the rapid rise of hyperscale data centers, synergy research group predicts that more than 120 to 130 hyperscale data centers costing billions of dollars will be launched each year in the future, and the cost of each data center will be in billions of dollars.

however, the high cost of ai and large models is not only reflected in the speed of burning money, but also in the reality that it is difficult to recover the investment in the short term. although openai's annual revenue has exceeded us$3.4 billion, high operating costs and fierce competition still keep it in a loss-making state. it is expected that by the end of 2024, the loss will be close to us$5 billion.

faced with this situation, domestic technology giants have shown a more cautious and pragmatic attitude, and have begun to shift from extensive exploration of general big models to deep-dive layout of industry vertical big models. as the cornerstone for coping with a wide range of tasks, the importance of general big models cannot be ignored; while vertical big models have become the new favorite of the market with their deep exploration and precise satisfaction of specific industries. in the fields of medical care, finance, law, education, etc., vertical big models are gradually showing huge application potential and reshaping the face of the industry.

it is worth emphasizing that the widespread application of large models cannot be separated from the solid support of computing power, data and algorithms, which sets a high threshold for small and medium-sized enterprises or enterprises with limited computing resources. however, it is precisely such challenges that have prompted the entire industry to continuously explore innovation and seek more efficient and economical solutions to promote the popularization and deepening of ai technology.

3

breakthrough: in china, ai is reshaping tens of millions of industries

the reason why vertical big models are popular is not only because of their relatively low development costs and controllable difficulty, but also because they can bring differentiated competitive advantages to enterprises. through intensive cultivation, enterprises can significantly improve the intelligence level of products and services, thereby standing out in the fierce market competition. however, the field of vertical big models is still a fertile ground to be developed. traditional enterprises often find it difficult to independently develop big models due to weak it foundations and input-output ratio considerations, and are more inclined to explore with the help of external forces.

even though vertical big models have lowered the threshold, they still face two major difficulties in the actual implementation process: one is how to find suitable application scenarios and design cost-effective product forms to achieve the best results at the best cost; the other is the compatibility issue between strategic planning and software and hardware facilities. the ambiguity of customer goals, insufficient technical knowledge and the complexity of system integration all increase the difficulty of implementing big models.

industry insiders pointed out that the goal of enterprises should be to use ai to solve practical problems, rather than simply pursuing integration with large models. therefore, enterprises need to think about how to better collaborate between people and machines, starting from problem solving, and avoid blindly pursuing large models. against this background, domestic leading technology companies are actively adjusting their strategic direction, shifting their focus to the refinement and deep cultivation of vertical large models, striving to maximize the actual application value of large models, and helping companies achieve a double leap in efficiency and benefits. this trend is particularly evident in china. among them, huawei cloud's technologies and products are helping all walks of life achieve intelligent transformation.

in the past, new drug research and development was a complex and lengthy systematic project. the pharmaceutical and health industry has long faced the dilemma of long new drug research and development cycles, high costs, and low success rates. under the traditional model, new drug research and development is often constrained by the "double ten rule", that is, ten years, one billion us dollars in investment, and a very low success rate. this not only restricts my country's pharmaceutical innovation, but also makes it difficult for many pharmaceutical companies to move forward on the road of research and development.

for the pharmaceutical and health industry, huawei cloud has launched a healthcare solution with pangu big model as the core, which improves the efficiency of drug design by 33%, and the optimized molecular binding energy by more than 40%, realizing the acceleration of the entire process in the early research stage, solving the problem of "innovative drugs", and making the "double ten law" no longer trouble drug research and development. based on pangu drug molecular big model, the first affiliated hospital of xi'an jiaotong university successfully reduced the drug research cycle from several years to several weeks, and reduced the capital and labor costs by 90%; southeast university used pangu drug molecular big model to achieve a result with a consistency of more than 80% with the experimental verification of organ chips; tasly used pangu drug molecular big model to learn 3.5 million natural product molecular data and created a "digital herbal big model", which improved the prediction and optimization results of natural product properties by 10%.

in order to open up the "last mile" of ai implementation, huawei cloud provides "ai+industry" problem-solving ideas that appear in more and more industries. for the implementation of large models in manufacturing scenarios, hailiang group and huawei have created the hailiang copper foil process optimization large model, which improves production efficiency through process optimization, improves product quality through formula optimization, and accelerates product innovation through process simulation. for the implementation of large models in the logistics industry, sf express, with the support of huawei cloud ascend cloud service, has created the "fengyu" large model and built a grand map of aigc applications. among them, huawei cloud ascend cloud service provides surging computing power support, and through efficient data, development, training and reasoning platforms, it helps the efficient development and resource utilization of ai applications, and builds a solid large model foundation.

in knowledge-intensive industries, the application of big models is driving the paradigm shift of "intelligence as a service". in the automotive industry, gac ai r&d assistant relies on huawei cloud codearts to build a code assistant and a diagnostic assistant, which realizes automatic code writing and detection, as well as automatic diagnosis of smart car faults. focusing on the problems of the difficulty of implementing and replicating artificial intelligence in the mining field, relying on huawei pangu mining big model, yunding technology has built an intelligent solution with a "1+4+n" architecture to create the first big model in the mining industry in china. through 1 ai development platform and 4 pangu big model capabilities, n high-value application scenarios are developed to support large-scale "going down the mine" of artificial intelligence.

today, the application of ai technology has penetrated into all walks of life and has become a new engine to promote global economic development. at the same time, it also faces multiple challenges such as scarcity of key data, professional scene recognition, and the difficulty of balancing model performance and cost-effectiveness. at the 2024 huawei all-connect conference, xu zhijun, huawei's vice chairman and rotating chairman, emphasized that ai technology is becoming the technology with the greatest impact on the industry. through ai, the industry is digitized and the industry's production methods are changed, becoming the core engine for various industries to enter the intelligent world. he proposed that enterprises in the era of intelligence should have "six a" characteristics, including adaptive experience, self-evolving products, autonomous operations, enhanced employees, full-volume, full-factor, full-connection, and intelligent native infrastructure.

as the world's second largest economy, china has abundant resources in application scenarios and data such as manufacturing, medical care, transportation, and home, providing a broad test field and application space for the development of ai. the development prospects of china's ai field are broad and full of opportunities. however, only when the industry jointly explores solutions can the coordinated development of technology and applications be truly promoted. in the face of the various challenges of the ai ​​era, only industrial integration and multi-party collaboration can achieve a win-win situation. by deeply exploring scenario applications, china will be able to continuously unleash the potential of ai technology and promote the digital transformation of various industries.

source: global times

process editor: u022