How DeepSeek re-engineers the future of AI applications in the real world

By Hong Kong Economic Times

Artificial intelligence (AI) has evolved from a concept to an indispensable tool in modern engineering. Powered by deep learning and multimodal reasoning, AI is driving transformative development across industries, such as using generative AI to optimise complex systems or making predictive analysis to enable proactive maintenance.

Against this backdrop, the open-source AI model built by Chinese AI company DeepSeek has captured global attention. After achieving performance comparable to top-tier models like GPT-4o and Claude Sonnet 3.5 with DeepSeek-V3 in December 2024, the latest DeepSeek-R1 has surpassed its predecessor, achieving GPT-o1 level performance at just one-tenth, or about US$5.6 million, the cost of similar models by its competitors. Additionally, it is an open-source model.

By rewriting the rules of AI innovation, DeepSeek adopts solutions such as Mixture of Experts (MoE) to reduce computational costs and boost efficiency. These strategies make it particularly well-suited for engineering applications, where real-world data variability and computational constraints are key challenges.

The rapid growth of AI offers huge opportunities for engineers. The demand for AI engineers is projected to grow at a compound annual growth rate (CAGR) of 20.17 percent, reaching a market size of US$9.46 million in 2029 from US$3.775 million in 2024, according to a report by Research and Markets¹. These professionals specialise in developing, programming, and training AI systems, with areas like Natural Language Processing (NLP), computer vision, machine learning, and robotics driving their roles. The rapid adoption of AI across industries, coupled with increased research and development investments and supportive government policies, has expanded their scope. As organisations are working to use AI to enhance efficiency, automation, and decision-making, the need for skilled AI engineers continues to rise, positioning them as key players in shaping future technological advancements.

The current state of AI in real-world applications

The integration of AI in real-world engineering has seen significant advancements, particularly in traditional applications such as computer-aided design (CAD)/ computer-aided engineering (CAE) automation and Internet of Things (IoT)-driven predictive maintenance. Leading platforms like Ansys and Siemens NX now integrate AI, including topology optimisation, to speed up design and automate processes. Below, we explore some examples of AI’s impact on different industries.

Transport: Volvo is accelerating its push into autonomous trucking through a partnership with AI startup Waabi. The partnership allows the integration of generative AI-powered self-driving technology into the Volvo VNL Autonomous truck that is manufactured in Virginia, the United States. By using Waabi’s AI virtual driver, these trucks can be trained to adapt to various scenarios that may be encountered on the road. The approach cuts down on resource usage, both human capital and financial investment, that is needed for traditional models. In terms of safety redundancies, the AI enhances Volvo’s existing six-layer safety system (covering steering, braking, power, and more) by predicting and mitigating risks that human engineers might overlook. Volvo has also introduced advanced AI for 3D simulations to streamline vehicle safety development and enhance efficiency in the manufacturing process². With the support of Nvidia’s AI supercomputing platform, the 3D rendering method Gaussian splatting allows the rapid generation of realistic, high-quality 3D scenes and objects derived from real-world imagery³.

Healthcare: AI is increasingly being used in the medical industry. For example, it improves resource management through predictive analytics, eases administrative workflows, and increases diagnostic accuracy via early detection of illnesses such as sepsis and breast cancer. This technology provides customised treatment strategies that reduce costs and improve patient outcomes. In the pharmaceutical sector, AI can play a role in drug development by refining clinical trials, and improving production standards. It also contributes to the monitoring of drug safety through data analysis⁴. In a report by the World Economic Forum in 2024, a case study on British digital patient platform Huma showed that it could reduce readmission rates by 30 percent, time spent reviewing patients by up to 40 percent, and ease the burden on medical staff⁵.

Financial services: The evolution of AI in financial services has been nothing short of revolutionary, with the technology being widely used in applications such as risk management, fraud detection, and customer service. Take JPMorgan Chase as an example, the US lender said that AI tools have helped its private client advisers find the right information faster by up to 95 percent⁶, while allowing its software engineers to increase their productivity by 10-20 percent with a coding assistant tool⁷. In Hong Kong, 75 percent of financial institutions have already implemented at least one Generative AI (GenAI) use case, or are currently piloting and designing use cases and exploring potential investment areas, according to a report by the Hong Kong Monetary Authority⁸ in April 2025. The ratio is expected to increase to 87 percent within the next three to five years.

Education: AI not only helps ease the burden of the teachers but also provides a personalised learning experience for students. From teaching, tutoring to exams, all aspects of the process can be transformed through AI technology to improve educational efficiency and teaching quality. In March 2024, The Hong Kong University of Science and Technology (HKUST) introduced AI lecturers through the utilisation of AI-generative tools. Developed using advanced AI tools—including 3D body generators, AI image generator Midjourney, ChatGPT, and face animation software, ten digital instructors were deployed to enhance engagement in various courses.

The HKUST created ten “AI Lecturers” from different origins, nationalities, professions and cultural backgrounds using AI-generative tools. (Photo downloaded from HKUST press release)

Technological innovations that position DeepSeek at the forefront of the AI revolution

In an AI race dominated by tech giants with billion-dollar budgets and cutting-edge hardware, DeepSeek has rewritten the rules of innovation. The AI pioneer has made breakthroughs in efficiency to deliver world-class performance while optimising every component of its architecture to come up with a powerful and cost-effective AI model.

The technological advantage of DeepSeek stems from three core innovations. First, the AI reasoning model of MoE, Mixure of Experts as previously mentioned. By only activating the necessary neurons for each task, this neural network design divides tasks among smaller expert networks, with only around 37 billion out of hundreds of billions of parameters activated per task. This helps significantly reduce computational costs and is crucial for making large-scale AI models more affordable. DeepSeek also enhances the MoE architecture through strategies such as Expert Choice routing to prevent knowledge overlap and balance workloads across experts of different fields⁹.

Second, DeepSeek introduced a novel attention mechanism called Multi-head Latent Attention (MLA) to optimise inference efficiency in Large Language Models (LLMs). Unlike traditional Multi-head Attention, which stores full key-value caches, MLA uses low-rank key-value joint compression to reduce memory overhead while maintaining a relatively high performance.

Last but not least, Grouped-query Randomised Parallel Optimisation (GRPO) is a critical component of DeepSeek’s technological advantage. It is a reinforcement learning (RL) algorithm specifically designed to enhance reasoning capabilities in LLMs. Compared with traditional approaches like Proximal Policy Optimisation (PPO), GRPO uses group-based relative comparisons instead of absolute rewards to remove the critic model. GRPO also enhances the stability of LLM training through relative ranking approach. Instead of relying on absolute reward scores, it evaluates responses by comparing them within groups, which reduces noise and improves learning efficiency. This allows DeepSeek to learn what may be the best strategies in complex scenarios¹⁰.

Apart from its technological advancement, DeepSeek’s strategy to open-source its models is key for its market expansion. This means all its implementation code and technical specifications are publicly accessible, allowing developers and institutions worldwide to use, adapt and implement. The strategy not only democratises AI technology but also encourages global developers to participate in improving and extending the model, which in turn drives rapid upgrade and technology sharing for a developers’ community and an AI ecosystem. Just as its founder Liang Wenfeng said, what is more important for the company is a strong technological ecosystem.

“Open-source and publishing research papers don’t really mean losing anything,” Liang said in an interview in July 2024. “For technical professionals, being followed and recognised is truly fulfilling.”

‘In fact, open-source is more of a cultural act than a business move. The act of giving is a kind of honour. When a company operates this way, it gains cultural appeal as well,’ said Liang¹¹.

As of the end of January 2025, DeepSeek overtook ChatGPT as the top free app on the Apple App Store in the US. Observers said that its rapid growth could be an essential step in democratising AI, allowing startups, independent developers, and smaller firms to utilise powerful AI capabilities through DeepSeek-R1. Some also believe that the shift could accelerate AI solutions globally, particularly benefiting regions with limited tech resources while fostering broader innovation¹².

Beyond traditional AI: DeepSeek in emerging fields

DeepSeek’s AI drives innovation in emerging sectors. Key collaboration areas include data engineering, autonomous vehicles, and human-AI decision systems. Below are some examples.

Climate and sustainability: In Shenzhen in South China’s Guangdong Province, the Municipal Ecology and Environment Bureau has incorporated DeepSeek-R1 model with environmental management in the Luohu district early this year. This marked the first application of AI technology in Luohu’s ecological and environmental sector. During the deployment, the team input key statistics such as PM2.5, PM10, SO2, and ozone, collected from monitoring stations across the district into DeepSeek’s model. Through in-depth data analysis done by AI-driven algorithms, the Bureau managed to optimise air pollution control strategies, pinpoint pollution sources for targeted emission reduction, identify potential environmental risks, and predict pollution dispersion trends to set up emergency planning in advance¹³.

In Chengdu, Sichuan Province, DeepSeek-R1 has been integrated into the water management platform, which acts as an advisor for environmental operations and provides expert solutions for issues like water treatment. The tool enhances management and decision-making efficiency.

The disruption by DeepSeek reflects China’s rapid advances in efficient AI development amid global tech competitions

Space and defense: In an interview with state-owned Chinanews.com, Wang Yongqing, chief designer at the Shenyang Aircraft Design Institute, said DeepSeek is being used to help develop China’s new-generation combat aircraft. The first aircraft design research institute in China has undertaken key projects for over 40 models, including J-8, J-11, J-15, J-35 and unmanned aircraft. With the rapid development of AI, the institute has introduced its first Digital Standard Review System and created digital staff to automate the standardised review of technical documents and design models, thus allowing researchers to focus more on critical scientific work. Media reports showed that DeepSeek has also been used in non-combat-related tasks within military settings, such as emergency evacuation planning and treatment plans for military doctors.

How DeepSeek is transforming engineering disciplines

A recent survey by Arup, the engineering consultancy, found that even the construction industry is making the use of AI a common practice, with about 36 percent of engineers, architects, and city planners relying on it daily. Furthermore, more than 80 percent are using advanced AI tools at least weekly. These AI-powered solutions are being used across workstreams including project design, advanced modelling, urban planning, creating digital twins, and enhancing sustainability and energy efficiency¹⁴.

In DeepSeek’s case, for example, designers at the Central South Architectural Design Institute Co Ltd., with the help of DeepSeek’s model, can quickly generate various style rendering after inputting keywords or uploading sketches. They can also adjust materials, lighting, and spatial proportions with the AI system, creating a new design norm of human-machine co-creation. The system is linked with the building information modelling (BIM) so that all data is seamlessly shared to enable the smart coordination of the entire “design-construction-operation” process¹⁵.

By providing real-time decision support, DeepSeek helps engineers respond swiftly to emergency issues, ensuring the safety of critical infrastructure. This capability is critical in fields such as civil engineering, where failure can result in severe consequences.

For electrical and computer engineers, DeepSeek’s advanced text recognition and visual reasoning capabilities can help them quickly identify circuit components and connections so that they can complete the design and troubleshoot issues more effectively. When urgent replacement is needed, engineers no longer need to read dozens of pages for alternative components because the system can automatically match key parameters such as on-resistance and gate charge and propose an optimal replacement plan.

In manufacturing, when DeepSeek-R1 detects minor glitches in the production line, engineers can use the AI tool to provide more innovative and timely solutions. The integration of DeepSeek with IoT devices can create a feedback loop where continuous data collection leads to more accurate predictions and recommendations. This can enhance the overall reliability and efficiency of manufacturing operations.

Global leadership and case studies: China’s AI ambitions and development in the West

In the United States, AI development is spearheaded by leading technology companies and research institutions. Companies like OpenAI, Google and Nvidia are at the forefront of AI research and development. According to the Stanford’s Artificial Intelligence Index, the US maintains its global leadership in state-of-the-art AI, producing 40 notable AI models in 2024. This nearly tripled China’s 15 and far surpassed Europe’s three.

According to Stanford, the US continues to dominate in private AI investment with US$109.1 billion invested last year, compared with China’s US$9.3 billion and the United Kingdom’s US$4.5 billion¹⁶.

The US also excels in the number of data centres, a key factor for AI development. According to Statista, the US had 5,426 data centres as of March 2025, followed by Germany’s 529, the UK’s 523. China had only 449 data centres¹⁷.

In spite of this, a report by US-based software company SAS Institute in 2024 showed that more organisations in China (83 percent) are using GenAI than those in the UK (70 percent), the US (65 percent) and Australia (63 percent). Notably, organisations in the US are ahead in terms of maturity and having fully implemented GenAI technologies than those in China and the UK.

“While China may lead in GenAI adoption rates, higher adoption doesn't necessarily equate to effective implementation or better returns,” said Stephen Saw, Managing Director at Coleman Parkes. “In fact, the US nudges ahead in the race with 24 percent of organisations having fully implemented GenAI compared to 19% in China.” ¹⁸

The disruption brought on by DeepSeek is another example of how China focuses on developing the technology “from one to ten” while the US excels in bringing it “from zero to one”. Though OpenAI first made the breakthrough in technology, DeepSeek demonstrates China’s ability to achieve high performance at a fraction of the cost compared to Western models through advancements in training techniques, leveraging open-source communities, and continued technological optimisation. It emphasises practical applications and cost-effective solutions that often build on existing technologies to create efficient products at a large scale.

In particular, the launch of a DeepSeek-powered AI service by the National Supercomputing Internet Platform in February 2025 has accelerated the application of DeepSeek models in various industrial scenarios and provided a more solid technological support for businesses. The newly launched service supports multiple versions of the DeepSeek-R1 model, including popular specifications from 1.5 B to 14 B. There are also plans to develop larger models of 671 B.

Needless to say, the AI competition between China and the West is reshaping the global AI landscape and pushing Western AI companies to re-evaluate their strategies. As a result, DeepSeek is not only expanding its market presence but also influencing the global AI industry to focus more on cost-effective and scalable solutions, which will eventually benefit industries and consumers worldwide.

The rise of DeepSeek also reflects China’s ambition in AI, which aims to become the premier AI innovation centre by 2030. In January 2025, the Ministry of Industry and Information Technology and the Ministry of Finance announced a state fund worth RMB 60 billion (US$8.2 billion) for early-stage investments in AI projects to strengthen the country’s AI industry¹⁹. Besides DeepSeek, leading tech companies such as Baidu, Alibaba and Tencent are investing heavily in AI.

Today, the competition in AI has gone beyond technology as it has emerged as a key battleground in global geopolitics. It is expected that both China and the US will continue to compete fiercely to ensure state power leadership, as well as economic and security dominance.

Challenges: engineering realworld AI systems

Engineers need to select appropriate resources such as Central Processing Units (CPUs), Graphics Processing Units (GPUs), Tensor Processing Units (TPUs), and servers for hardware while integrating them with proper software solutions to ensure a solid AI infrastructure layer.

But the complexity and diversity of engineering problems in AI and machine learning pose massive challenges for engineers.

These include data quality and scalability, computational constraints, interpretability, real-world deployment, and edge AI limitations.

For example, the quality of data serves as the foundation for the effectiveness of AI models. Yet missing values, inconsistencies and noise are often among the issues that affect the reliability of the model. In terms of safety and regulatory requirements, AI systems face stringent safety standards and verification process, particularly for high-risk sectors such as defense, aerospace and healthcare. Limitations in computing resources and deployment bring challenges to model optimisation.

The “black boxes” nature of AI, such as complex models like deep neural networks, also makes it difficult to clearly understand how the input data is transformed into the output. The lack of transparency increases the difficulty for engineers to debug the system or locate the problem. It will also affect users’ trust in AI systems, especially in areas that involve high-risk decisions, such as autonomous driving and financial management. Therefore, improving the transparency and interpretability of the models becomes key to understanding how the predictions and decisions are made.

The future: AI-augmented engineering for AI-powered future

The impact of AI on global development is evident. In the future, engineers will continue to play a key role in driving AI-related innovation. By integrating AI into different engineering disciplines, engineers can help realise automated design, intelligent optimisation and autonomous decision-making. In product development, the use of GenAI can quickly generate design solutions. In manufacturing, an AI-powered predictive maintenance system can help improve equipment reliability. In infrastructure and construction, AI can help civil engineers better manage resources and energy consumption. The human-machine collaboration supported by AI will help engineers and users to better face challenges in the era of technology and innovation.

Engineers are also responsible for developing AI technologies in an ethical way, so as to make them transparent, accountable and unbiased. When integrating AI into engineering applications, engineers need to consider the technology’s data privacy, algorithmic fairness, and social impacts. They should also work with policymakers to establish proper guidelines in AI technology to ensure the innovation is on the right path.

The adoption of AI in various sectors promotes the development of human-machine collaboration.

References

1. Research and Markets (2024). AI Engineers Demand Outlook Report 2024-2029: Adoption of Cutting-Edge Technologies in NLP, Robotics, and Computer Vision at the Forefront. GlobeNewswire News Room. Available at: https://www.globenewswire.com/news-relea se/2024/11/15/2981824/28124/en/AI-Engineers-Demand-Outlook- Report-2024-2029-Adoption-of-Cutting-Edge-Technologies-in-NLPRobotics- and-Computer-Vision-at-the-Forefront.html [Accessed 1 May 2025].
2. Chapman T (2025). Volvo and Waabi: Advancing Autonomous Trucks in Logistics . Supplychaindigital.com. Available at: https://supplychaindigital.com/logistics/how-volvos-autonomous-trucksare- gen-ai [Accessed 1 May 2025].
3. Brown R (2025). How Volvo’s AI 3D Worlds are Enhancing Manufacturing Safety . Manufacturingdigital.com. Available at: https://manufacturingdigital.com/ai-and-automation/inside-volvosinnovative- ai-powered-safety-programme [Accessed 1 May 2025].
4. European Commission (2025). Artificial Intelligence in healthcare . Public Health. Available at: https://health.ec.europa.eu/ehealth-digital-health-and-care/artificialintelligence- healthcare_en [Accessed 1 May 2025].
5. North M (2025). 6 ways AI is transforming healthcare. World Economic Forum. Available at: https://www.weforum.org/stories/2025/03/ai-transforming-globalhealth/ [Accessed 1 May 2025].
6. Anand N (2025). JPMorgan says AI helped boost sales, add clients in market turmoil . Reuters. Available at: https://www.reuters.com/business/finance/jpmorgan-says-ai-helpedboost- sales-add-clients-market-turmoil-2025-05-05/ [Accessed 5 May 2025].
7. Suresh H (2025). JPMorgan engineers’ efficiency jumps as much as 20% from using coding assistant . Reuters. Available at: https://www.reuters.com/technology/artificial-intelligence/jpmorganengineers- efficiency-jumps-much-20-using-coding-assistant-2025-03-13/ [Accessed 5 May 2025].
8. Hong Kong Monetary Authority (2025). Hong Kong Monetary Authority - Report on ‘Financial Services in the Era of Generative AI: Facilitating Responsible Adoption’. Available at: https://www.hkma.gov.hk/eng/news-and-media/pressreleases/ 2025/04/20250409-3/ [Accessed 5 May 2025].
9. Sayed Ali Alkamel (2025). DeepSeek and the Power of Mixture of Experts (MoE). DEV Community. Available at: https://dev.to/sayed_ali_alkamel/deepseek-and-the-power-of-mixtureof- experts-moe-ham [Accessed 5 May 2025].
10. Ahmed S (2025). The Math Behind DeepSeek: A Deep Dive into Group Relative Policy Optimization (GRPO) . Medium. Available at: https://medium.com/@sahin.samia/the-math-behind-deepseek-adeep- dive-into-group-relative-policy-optimization-grpo-8a75007491ba [Accessed 5 May 2025].
11. 暗涌Waves. ‘對話幻方梁文鋒：AI界的拼多多——揭秘DeepSeek！.’ Wallstreetcn.com. https://wallstreetcn.com/articles/3719982. [Accessed 5 May 2025].
12. Edmond C (2025). What is open-source AI and how could DeepSeek change the industry? World Economic Forum. Available at: https://www.weforum.org/stories/2025/02/open-source-ai-innovationdeepseek/ [Accessed 5 May 2025].
13. 深圳市生態環境局 (2025). 深圳市生態環境局羅湖管理局創新應用DeepSeek大 模型賦能區域環境治理. Available at: https://meeb.sz.gov.cn/fzlm/swrfzf/content/post_12061000.html [Accessed 5 May 2025].
14. Allen D (2025). Architects and engineers relying daily on AI to design cities and infrastructure, global survey reveals . Arup.com. Available at: https://www.arup.com/news/architects-and-engineers-relying-daily-onai- to-design-cities-and-infrastructure-global-survey-reveals/ [Accessed 7 May 2025].
15. Archina.com. (2025). AI賦能 | 中南建院‘GIANT AI’全面接入 DeepSeek —— 讓AI成為設計師的‘超腦搭檔’ ARCHINA 資訊. Available at: http://www.archina.com/index.php?g=portal&m=index&a=show& id=10066 [Accessed 7 May 2025].
16. Fried I (2025). U.S.’s AI lead over China rapidly shrinking: Stanford report . Axios. Available at: https://www.axios.com/2025/04/07/china-ai-race-stanford-report [Accessed 7 May 2025].
17. Taylor P (2024). Data Centers Worldwide by Country 2021. [online] Statista. Available at: https://www.statista.com/statistics/1228433/data-centers-worldwide-bycountry/ [Accessed 7 May 2025].
18. NC C (2024). Global Market Research: China leads world in GenAI usage while US leads in full implementation . www.sas.com. Available at: https://www.sas.com/en_us/news/press-releases/2024/july/genairesearch- study-global.html [Accessed 14 May 2025].
19. Jiang B (2025). Tech war: China creates US$8.2 billion AI investment fund amid tightened US trade controls . South China Morning Post. Available at: https://www.scmp.com/tech/big-tech/article/3295513/tech-warchina- creates-us82-billion-ai-investment-fund-amid-tightened-us-tradecontrols [Accessed 14 May 2025].