Why DeepSeek chose text over images: The untold strategy behind its AI success
DeepSeek’s focus on cost-effective text processing has captivated the AI industry. As both China and the US compete and cooperate, the future of AI innovation unfolds. Technology expert Yin Ruizhi explores DeepSeek’s strategic edge for winning the race.
DeepSeek’s rapid growth and global attention raise an interesting question: why has it captivated the AI industry in both China and the US, and inspired numerous emerging enterprises to build their businesses around it?
The answer lies in its unwavering focus on cost reduction and text processing, rather than pursuing features like text-to-image generation, which has been OpenAI’s claim to fame. I have already highlighted the fact that DeepSeek’s low cost has attracted widespread attention within the industry in another article of this series published in 2024.
As early as a year ago, a consensus emerged among leading Chinese AI professionals that the first area where this wave of large language model (LLM) technology would deliver stable commercial value is high-quality text analysis, summarisation and review, rather than content generation. Hence, the term “generative AI”, frequently used in the media, is quite misleading to those outside the AI field.
When users view AI primarily from a generative angle, they might miss the important skills of judgment and intervention necessary for evaluating the generated content.
What is needed for AI large models to work well?
A key feature of this generation of large language models (LLMs) is that producing high-quality results often requires users to have a strong understanding of the results and the related issues. LLMs can sometimes “hallucinate”, meaning they generate content that is not accurate. When this happens, users need to step in to guide and correct the output, refining it through multiple iterations to achieve the best results.
When users view AI primarily from a generative angle, they might miss the important skills of judgment and intervention necessary for evaluating the generated content. This gap is why self-media often sensationalises AI’s impressive outputs, while the average user experiences only mediocre results in areas like music or video creation.
The key issue is users’ inability to effectively intervene and refine AI output. When tasks involve text analysis, summarisation and review, and the user possesses strong content judgment, AI can be readily applied. For instance, a teacher can utilise AI to review a large number of student assignments, guiding it to analyse details and assess students’ understanding of key concepts. Similarly, a financial analyst can leverage AI to process numerous market research reports, efficiently extracting valuable insights.
What industries could use AI LLMs?
China’s e-commerce sector is a beneficiary of this current wave of LLMs. E-commerce product owners need to process vast amounts of information, including extensive user feedback, competitor sales data, and other businesses’ marketing strategies, then analyse and summarise them, in order to formulate their own product strategies.
A product owner for a cookie business, for example, faces competition from over 37,000 similar products on mainstream e-commerce platforms like Taobao, JD.com and Pinduoduo. Prior to the existence of LLMs, this market analysis work typically required an entire team. Beyond personally reviewing a substantial amount of information, product owners relied on their teams to summarise information they were unable to process themselves. Now this presents a problem: if a member of this team slacks off or makes a mistake, the product owner may miss important information.
In the food and toy e-commerce projects that I was a part of, such errors could directly lead to the failure of marketing plans and product upgrade plans, with more than a 40% chance of this happening. With the advent of large AI models, a product manager no longer needs an assistant. They simply need to collect the target data and have the large model “faithfully” read through it, providing feedback according to the user’s requirements.
Presently, China’s e-commerce sector is also the most proactive and effective in applying this round of large AI model technology.
Since the product manager is an expert in the field, they can more easily discern the results, while at the same time instruct the AI to continually refine its output, or even directly read the source data. Thus, in such application scenarios, large AI models can significantly improve efficiency without letting shortcomings such as LLM hallucinations affect the outcome.
One key reason DeepSeek initially prioritised cost reduction over releasing features like text-to-image generation is that the first wave of real-world applications for large models was in content analysis, summarisation, and textual auditing — a field where cost sensitivity is crucial.
China is currently the most developed country globally when it comes to the e-commerce industry, bar none (even the US’s e-commerce industry is not as advanced as China’s). Presently, China’s e-commerce sector is also the most proactive and effective in applying this round of large AI model technology.
In the true industrial domain for AI, there is a lot of exchange and integration between China and the US in terms of talent and technology — the information gap is small.
Does the US know about all this?
Another interesting question is whether US counterparts are aware of this fact. The answer is: they are. In the true industrial domain for AI, there is a lot of exchange and integration between China and the US in terms of talent and technology — the information gap is small.
Two points can prove this. Following the US election, Elon Musk, the richest man in the world who is in charge of the Department of Government Efficiency, used AI technology to extensively audit the expenditures of various US government departments, identifying areas of unreasonable expenditure. The direction of such an application is essentially for the purpose of high-quality analysis, summarisation and auditing of large-scale information.
At the same time, OpenAI is also very aware of this direction. Following the rousing success of DeepSeek, among the responses from OpenAI, the release of Deep Research was most noteworthy to those in the industry. It provides in-depth analysis reports for professionals in fields such as finance, science and engineering; for example, financial analysts might use it to analyse industry supply chain risks.
In the Humanity’s Last Exam test, the model used by Deep Research achieved an accuracy of 26.6% on expert-level questions, breaking previous records. It also performed well in the GAIA evaluation, achieving high accuracy across various difficulty levels, especially in level 3 tasks that require complex multi-step research and synthesis. This application direction is also essentially high-quality analysis, summarisation and auditing of large-scale information.
The mutual competition and co-prosperity of China and the US in the AI field is only just beginning.
