Deepseek vs. Openaai: Ki-Wet racen exposes-is China's R1 just a copy or a strategy masterpiece?

Technology race of the giants: DeepSeek vs. OpenAI – Who will dominate the future of AI?

China and the US have been at the heart of global technological development for years. Particularly in the field of artificial intelligence (AI), an intense race is underway, with large tech companies and emerging startups alike searching for innovative solutions. In this context, the Chinese AI startup DeepSeek and the American company OpenAI have attracted considerable attention. DeepSeek recently unveiled two remarkable AI models: DeepSeek R1 (the basic version is called "R1") and DeepSeek R1 Zero (often also referred to as "R1-Zero"), while the US company OpenAI has presented its o1 model and its smaller variant, o1 mini. Many observers are wondering whether the DeepSeek R1 and R1 Zero models are merely accidental imitations of US technologies or whether they represent a deliberate strategy aimed at propelling the Chinese AI sector to prominence.

This text delves into the differences and similarities between the AI ​​systems of DeepSeek and OpenAI. Furthermore, it examines how reinforcement learning is applied in DeepSeek R1 Zero and R1 and explores the potential benefits for next-generation AI models. This discussion will comprise over 2,000 words, enabling a comprehensive overview and in-depth analysis. At the same time, it strives to present only reliable information, avoiding pure speculation and focusing instead on verifiable trends, established technical data, and statements from the AI ​​field.

Suitable for:

• China vs. USA in AI: Are DeepSeek R1 (R1 Zero) and OpenAI o1 (o1 mini) really that different?

Global competition in the AI ​​sector

Competition between China and the US in the field of AI has intensified significantly in recent years. Observers frequently describe the two countries as being in a veritable race for dominance in the future technology of AI. This intensification of the competition stems from several factors. First, policymakers in both nations see AI as having the potential to secure innovation leadership for decades to come. Second, major technology companies have recognized the enormous economic benefits that AI solutions promise. Third, both China and the US have formulated comprehensive strategies to advance AI research.

In China, AI has been considered a key component of the country's modernization and a "key to international competitiveness" for several years. The government supports startups and research institutions with diverse programs and funding to expand the development of AI technologies. In contrast, the US relies on the power of the free market, where large and established companies like Google, Microsoft, Meta, and OpenAI, as well as many smaller players, compete and receive substantial funding from investors to advance machine learning, neural networks, and natural language processing (NLP).

DeepSeek and OpenAI at a glance

As an emerging player from China, DeepSeek has become something of a "hidden gem" in the global AI scene. The AI ​​startup is less well-known than the major Chinese tech companies, but has garnered attention in expert circles because it appears to be developing high-quality Large Language Models (LLMs) in a short time. Two of these models are DeepSeek R1 and DeepSeek R1 Zero. OpenAI, on the other hand, is a California-based company that is globally renowned for its AI models and gained early recognition. With o1 and its smaller sibling, o1 mini, OpenAI demonstrates its focus on high-quality and scalable AI systems.

The DeepSeek R1 and R1 Zero models recently achieved benchmark results comparable to OpenAI's o1 mini and the more powerful o1 model. In an industry where innovation is often dominated by well-known US corporations, the Chinese company DeepSeek has suddenly become a serious competitor. Some analysts are questioning to what extent DeepSeek was inspired by US approaches and whether it merely copied strategies or actually introduced new ways of thinking.

Technical basics of DeepSeek R1 and R1 Zero

1. DeepSeek-R1-Zero: Reinforcement Learning without human supervision

DeepSeek-R1-Zero is attracting particular attention because this model relies entirely on reinforcement learning (RL) without prior human feedback or traditional supervised fine-tuning. This approach is considered remarkable, as the majority of advanced AI applications rely on human-annotated data or feedback from real-world testing, at least in some phases.

DeepSeek-R1-Zero takes a different approach. The model was designed to develop the ability to recognize large and complex relationships and to improve independently. Through the consistent use of real-life feedback, R1-Zero has acquired specific skills that are particularly relevant in the area of ​​reasoning. These include:

• Self-checking: Before giving a final answer, the model checks its own intermediate steps (its “inner monologue”) to uncover errors.
• Reflection: Instead of directly providing a single answer, the model reflects on different answer options, similar to how a person weighs possible solutions against each other.
• Generating long chains of thought: R1-Zero shows that it can generate intermediate steps even for complex tasks, which it uses flexibly in the solution.

The ability to self-monitor and restart when encountering a dead end is considered crucial for future breakthroughs in AI research. The more complex the problem, the more important the capacity to organize thought processes and correct flawed approaches becomes.

2. DeepSeek-R1: Combination of reinforcement learning and classic fine-tuning

The sister model DeepSeek-R1 combines the potential of reinforcement learning with the more traditional approach of supervised fine-tuning. The rationale behind this strategy is that while reinforcement learning can lead to particularly creative and elegant solutions, it sometimes misses the mark when it comes to human expectations regarding comprehensibility and relevance. To counteract this, DeepSeek's developers have additionally implemented fine-tuning methods that utilize human feedback and curated training data.

According to internal tests and several publicly available benchmarks, DeepSeek-R1 demonstrates strong performance in various areas. These include:

• Mathematics: Average accuracy of 79.8% for AIME and 97.3% for MATH-500.
• Programming: In code competitions like Codeforces, the model outperforms approximately 96.3% of other participants.
• General knowledge: DeepSeek-R1 shines here with a score of 90.8% for MMLU and 71.5% for GPQA Diamond.

The fact that DeepSeek-R1 is more cost-effective yet achieves excellent results in many areas has piqued the interest of observers. "Is this the beginning of a new AI era in which startups challenge the highly funded US giants?" some commentators are asking.

OpenAI's o1: Background, Philosophy and Achievements

From the outset, OpenAI has strived to develop “safe and useful AI for the benefit of humanity.” This guiding principle is reflected in many decisions, including the combination of reinforcement learning and human feedback (RLHF). The idea behind this is that the model learns through interaction with human feedback providers to give answers that are not only formally correct, but also understandable, helpful, and ethically sound to humans.

RLHF aims to prevent potential problems, such as a model generating inappropriate content. However, this requires additional resources, as maintaining and training the model, including human review and feedback processes, is costly. These costs are often reflected in higher subscription or usage fees. For example, o1 is frequently criticized for its comparatively high API prices, while other providers, such as DeepSeek, offer lower barriers to entry.

In terms of performance testing, OpenAI's o1 is considered a powerful system applicable to a wide range of tasks. From mathematics and programming to creative text generation, o1 has repeatedly demonstrated its high level of performance. Its Chain-of-Thought Reasoning is particularly well-known, as the model breaks down complex problems into intermediate steps and delivers highly precise results. For example, someone posing a mathematical word problem can often follow the thought process. While the model doesn't reveal every single step transparently, it usually provides a step-by-step argument that leads to a clearly comprehensible solution.

Comparison of the two systems: DeepSeek-R1 vs. o1

1. Performance differences

Mathematics tests reported that DeepSeek-R1 achieved an accuracy of 79.8% for AIME, while o1 reportedly reached 79.2%. This is a minimal difference, but it has a psychological impact because DeepSeek presents a technically equivalent or even slightly superior model. In programming, DeepSeek-R1 reportedly achieved around 96.3% in the Codeforces test, while o1 was said to be just over 96.6%. This difference is also small, but it shows that both models are performing at a comparable level.

2. Costs and accessibility

A key point is the differing cost structure. While OpenAI charges relatively high fees for o1, DeepSeek-R1 reportedly operates at significantly lower prices: "Up to 95% cheaper" is stated in some DeepSeek company presentations. Such claims need to be verified in practice, but if this cost advantage proves true, it could prove to be a major competitive advantage for DeepSeek. This is especially true for enterprise customers who need to process enormous amounts of data and therefore opt for a solution that saves costs in the long run.

Furthermore, DeepSeek-R1 is available under the MIT license, which allows the free use and modification of the model weights and outputs. In a time when more and more developers and companies are relying on open source, this could be a decisive advantage. "For us, openness means fostering innovation" is a statement that DeepSeek repeatedly emphasizes. Open-source solutions allow developers to directly access the code, make adjustments, and integrate the model into their own projects without being forced into a closed ecosystem.

Suitable for:

• AI models in numbers: Top 15 major language models – 149 foundation models – 51 machine learning models
• AI language models BERT and GPT: Which companies are behind them, what exactly can they do, and what are the differences?

3. Special abilities

Both DeepSeek-R1 and o1 are characterized by advanced reasoning. DeepSeek-R1, through RL (Reference-Based Reasoning), developed a pronounced capacity for self-critical reflection, coordinating intermediate thought processes and "long chains of thought." OpenAI's o1, on the other hand, excels in chain-of-thought reasoning, which refers to the ability to create step-by-step and logically traceable solution paths. Both models are therefore able not only to present results immediately but also to explain their reasoning to a certain extent. This increases transparency and confidence in the output.

DeepSeek-R1 Zero: Specializations and Outlook

1. Focus on Reinforcement Learning

DeepSeek-R1 Zero is, in a sense, the radical version of the R1 model, as it foregoes traditional human feedback. While R1 relies partly on supervised fine-tuning, R1-Zero depends entirely on real-world feedback. From an AI research perspective, this is an exciting experiment: "The potential of reinforcement learning is being pushed to its absolute limit here," some observers say. Reinforcement learning mimics the principle of trial and error, in which the model receives reward signals for correct intermediate steps or final results.

A key element of R1-Zero is its ability to take its time to think. If a particular problem is deemed more difficult, the model uses more computation cycles to search for a suitable solution. While this adaptive compute approach can slow down the model's response, it tends to improve the quality of the results. "Slower, but smarter" is a fitting summary.

2. Challenges

However, the radical reinforcement learning approach also has its downsides. DeepSeek-R1 Zero is said to sometimes suddenly switch between different languages ​​or generate output that is confusing from the user's perspective. This uncontrolled language switching could be due to variant exploration phases in the reinforcement learning process. Furthermore, it remains unclear how the reinforcement learning methodology will perform in real-world application scenarios in the long term, where fault tolerance is sometimes lower and regulatory requirements are high.

R1-Zero currently lacks advanced dialog functions, JSON output, and specialized function calling. Such features are often essential for integrating AI solutions into business environments, for example, for automated processes. DeepSeek has announced plans to gradually add these functionalities. However, it remains to be seen whether and when these updates will be released.

Democratizing AI through open source?

DeepSeek has not only released its large models R1 and R1-Zero, but is also making six smaller derivatives publicly available. These models were partially trained using data extracted from the larger models. The goal is to provide AI developers worldwide with easy-to-use tools to build their own AI projects. "We want the AI ​​revolution to reach everyone, not just large companies or research institutes," DeepSeek stated.

Such steps could truly transform the AI ​​landscape. If powerful models are openly available, startups and independent developers won't need to enter into costly licensing agreements with large US providers; instead, they can directly modify and deploy their own versions of DeepSeek's models. Some experts see this as an opportunity to foster genuine diversity and innovation in AI by preventing monopolies and oligopolies.

Is it imitation or strategic in-house development?

A recurring theme in the East-West AI competition is: Is China simply copying approaches from the US, or is it genuinely developing its own? Indeed, DeepSeek R1 and R1 Zero show many parallels to the workings of OpenAI's o1. For example, both use reinforcement learning for process optimization. The idea of ​​incorporating a chain of thought into the logical processing of multi-step tasks also emerged early in Western research. Therefore, it's reasonable to assume that DeepSeek has also benefited from these insights and is implementing a similar paradigm in some respects.

However, such similarities should not be hastily interpreted as evidence of plagiarism or mere imitation. Research and development in AI is a globally driven field where new ideas spread rapidly. Furthermore, scientific publications deepen progress across the entire field, allowing researchers worldwide to build upon the same foundation. It could just as easily be that DeepSeek has independently refined its reinforcement learning approach to a point that, in some benchmarks, even surpasses its competitors.

Competitive opportunities and risks

Due to their impressive performance, DeepSeek R1 and R1-Zero are attracting interest from investors, research institutions, and technology companies. Anyone looking for a cost-effective, high-performance, and open solution can hardly ignore DeepSeek. "There aren't many providers that offer such a high level of performance while also providing this degree of openness," is the general consensus among some industry experts.

Nevertheless, risks remain. Some potential customers are hesitant to adopt "version 1" models, as AI systems often only reach market maturity after several iterations. Furthermore, it is unclear whether DeepSeek can guarantee the necessary stability and reliability in its support processes, which are crucial for large clients. Questions regarding warranties, trustworthiness, data protection, and security are also essential. Especially when dealing with sensitive data, not only is technical performance decisive, but also whether the AI ​​solution meets the security requirements of international companies.

Ethical and geopolitical implications

Geopolitical tensions between China and the US in the technology sector are increasingly being projected onto the AI ​​sector. Many companies are asking themselves, "Who can we trust when it comes to sensitive data and the development of novel AI agents?" On the Western side, there is skepticism towards Chinese AI systems due to fears of potential interference by government agencies. Conversely, there are reservations in China about US dominance and potential backdoors in proprietary systems.

This conflict is reflected in the question of whether DeepSeek truly represents an independent innovation or is merely a copy “made in China.” If it could be demonstrated that DeepSeek R1 and R1-Zero set new quality standards, China would possess one of the leading AI systems, which, from a geopolitical perspective, would symbolize the country's rapid technological rise. Conversely, the success of OpenAI's o1 and its continued development in the US could ensure that American AI companies retain their dominance in shaping the market.

Potential application scenarios

1. Scientific research and mathematics

Both DeepSeek-R1 and o1 are of interest to researchers, students, and educational institutions due to their strong performance in mathematical problems. Thanks to high accuracy scores in areas such as AIME and MATH-500, these models are suitable for solving complex algebraic, geometric, and analytical problems. They can also serve as tools for extracting and summarizing scientific texts.

2. Programming and Software Development

These models could also prove useful in software engineering. DeepSeek-R1 and o1 can interpret source code, identify faulty sections, and suggest optimizations. DeepSeek-R1 also integrates a feature that allows code to be tested and rendered directly within a chat interface. This accelerates development cycles and promotes rapid iterations. Developers working in teams could thus benefit from a virtual code coach who provides continuous feedback.

3. Creative brainstorming and content creation

Both models can support text creation processes by generating ideas, suggesting content structures, or assisting in the writing of longer articles. This opens up new possibilities for copywriters, journalists, and bloggers to create content efficiently and continually introduce fresh perspectives. However, it remains crucial to critically evaluate the output and not adopt it blindly.

Looking to the future: Will DeepSeek and OpenAI shape the AI ​​market?

The further development of DeepSeek R1 and R1-Zero could signal a global trend toward powerful, autonomous AI models that learn independently and require only limited human intervention. The increased focus on reinforcement learning reflects a general direction in modern AI research. Once these models prove their worth in real-world projects, other companies are likely to follow suit.

OpenAI, for its part, will strive to maintain or even extend its lead. The company is researching further developed versions of o1, which promise even more precise chain-of-thought capabilities, improved dialogue interfaces, and stronger security mechanisms. Cost reduction is also likely to play a role in the future, as more and more competitors enter the market.

Suitable for:

• AI & Metaverse: The dream team of the digital future – The next level of digitalization

A tension between innovation and competition

No, DeepSeek, with its R1 and R1-Zero models, is not simply a copy of US technologies, but rather possesses its own strengths and approaches. The assumption of strategic imitation cannot be entirely dismissed, as research findings in the AI ​​world are typically shared openly, and every player strives to adopt the latest methods. However, it would be an oversimplification to reduce DeepSeek to the label of "plagiarism." The benchmark results presented and the openness of the AI ​​models tell a different story.

“We are at the beginning of a new phase of the AI ​​revolution” is a frequently heard statement in Silicon Valley as well as in Chinese innovation centers. This statement sounds general, but it reflects a genuine paradigm shift: In this revolution, it is no longer just the big names that set the pace, but also a multitude of startups and research teams that are transforming the market with innovative ideas and affordable solutions. DeepSeek R1 and R1 Zero are an example of this that can no longer be ignored.

Of course, the question remains open as to which model will ultimately prevail, or whether both (and other competing products) will complement each other to form a global AI ecosystem. A coexistence where developers have the choice of implementing their projects with either US or Chinese models (or even a combination) would be beneficial for the overall culture of innovation. In any case, the technical soundness and reliability of the models remain crucial.

One thing is already certain: DeepSeek R1 and R1 Zero could help democratize AI by making advanced models accessible to a wider audience. If DeepSeek proves to be a high-quality yet cost-effective solution, the pressure on other vendors to redesign their pricing models or become more transparent will increase. OpenAI's o1, on the other hand, is considered by many to be the "gold standard" in terms of quality, stability, and community support. However, critics have also voiced their concerns, arguing that OpenAI's solutions are not affordable or flexible enough for every use case.

“Is it coincidence or strategic imitation in AI development?” – This question probably cannot be definitively answered. It is far more likely that DeepSeek and OpenAI each build upon a shared foundation of knowledge and draw inspiration from similar research findings. Both contribute their own ideas and innovations and strive to surpass the competitor in specific disciplines. In the long run, this competition can benefit everyone because it raises standards, accelerates technological progress, and reduces the cost of using AI-based services.

The AI ​​race between China and the US will continue, and with it the question of how established industry players will fare compared to emerging newcomers. There is very likely no simple answer to who will dominate in ten years. Too many factors – from geopolitical developments and the economic situation to cultural aspects – influence the overall technological landscape. What is an ambitious start-up today could be a leading global player in AI tomorrow; what is considered a leader today may have to contend with strong challengers tomorrow.

One thing is certain: Reinforcement learning, open licenses, fair pricing structures, and the ability to transparently map complex thought processes are key drivers of success and innovation. Companies that combine these factors while simultaneously ensuring the security and protection of sensitive data are well-received by the market. DeepSeek R1, R1 Zero, and OpenAI's o1 are excellent examples demonstrating that the time has come for a new chapter in AI. The world can look forward with anticipation to the further advancements that the next year and the coming decades will bring—and whether a new generation of LLMs will succeed in realizing the vision of truly universal AI.

This concludes our discussion of DeepSeek R1, R1 Zero, and their comparison with OpenAI o1. We see that the AI ​​landscape is constantly evolving, with new models constantly competing with established ones. This development is characterized by intensive research, mutual inspiration, healthy competition, and ever-greater challenges that must be tackled together. As these technologies advance, it will be increasingly interesting to see whether and how China and the USA combine their respective strengths or play them off against each other. Ultimately, society as a whole could be the winner if models like DeepSeek R1, R1 Zero, and o1 deliver innovative solutions that revolutionize how people process information, solve problems, and become creative.

From local to global: SMEs conquer the global market with clever strategies - Image: Xpert.Digital

At a time when a company's digital presence determines its success, the challenge is how to make this presence authentic, individual and far-reaching. Xpert.Digital offers an innovative solution that positions itself as an intersection between an industry hub, a blog and a brand ambassador. It combines the advantages of communication and sales channels in a single platform and enables publication in 18 different languages. The cooperation with partner portals and the possibility of publishing articles on Google News and a press distribution list with around 8,000 journalists and readers maximize the reach and visibility of the content. This represents an essential factor in external sales & marketing (SMarketing).

More about it here:

• Authentic. Individually. Global: The Xpert.Digital strategy for your company

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitalization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development

Konrad Wolfenstein

I would be happy to serve as your personal advisor.

You can contact me by filling out the contact form below or simply call me on +49 7348 4088 965 (Munich) .

I'm looking forward to our joint project.

Xpert.Digital is a hub for industry with a focus on digitalization, mechanical engineering, logistics/intralogistics and photovoltaics.

With our 360° business development solution, we support well-known companies from new business to after sales.

Market intelligence, smarketing, marketing automation, content development, PR, mail campaigns, personalized social media and lead nurturing are part of our digital tools.

You can find out more at: www.xpert.digital - www.xpert.solar - www.xpert.plus