1. [industry pattern] will DeepSeek's low-cost training model change the AI industry pattern? Will this lower the threshold for AI entrepreneurship, intensify the "Matthew effect" and affect the AI chip market?

Transformation of industry pattern

• the training cost is greatly reduced: the training cost of DeepSeek is only US $5.57 million, using 2048 H800 GPUs, which is more than 90% lower than that of traditional models.
• Lower access threshold: the open source strategy enables more developers to participate in and innovate, lowering the industry access threshold.
• The open source strategy further lowers the industry access threshold.

This low-cost and efficient training mode is breaking the resource barriers of traditional AI development and pushing the industry to develop in a more open and efficient direction.

Substantial reduction of entrepreneurial threshold

• funding threshold: significantly reduce initial investment, enabling small teams to develop high-performance models.
• Technical threshold: open source strategy allows more developers to participate and innovate.
• Infrastructure threshold: reduce dependence on high-end hardware and reduce operating costs.

This change will lead to a new wave of AI entrepreneurship.

Bidirectional effects of Matthew effect

• reduce barriers: allow more players to enter the market and reduce resource monopoly.
• Accelerated differentiation: enterprises with data and application scenario advantages may achieve faster development.

It is expected that there will be a new pattern of "The strong are stronger and the innovators coexist.

Impact on AI chip market

• in the short term: it may reduce the demand for high-end GPU, causing certain impact on chip manufacturers such as Avida.
• In the long run, the popularization and development of AI applications will bring greater chip demand. DeepSeek's technological innovation may also prompt chip manufacturers to develop more efficient and targeted AI chips.

Therefore, the emergence of DeepSeek will not completely subvert the AI chip market, but will promote its transformation and development.

2. [international competition] will the rise of DeepSeek have a significant impact on Sino-US AI competition? Will this affect export control policies, competition situation, and promote global AI strategy adjustment?

Impact on AI Competition Pattern

• DeepSeek trained a model comparable to GPT-4 at a low cost of 5.6 million US dollars, challenging the traditional American cognition that "high investment can achieve high performance.

• Using innovative reasoning time calculation method, the effect that requires 16000 NVIDIA chips can be achieved with 2,000 NVIDIA chips.

• The open-source strategy has enabled its technology to spread rapidly, and its applications have reached the top of the AppStore, shaking the dominant market position of American AI enterprises.

Dual impact on export control policy

• the United States may expand its control scope to cover more chip models (such as H20).
• Consider vulnerabilities such as restricting remote access to cloud services.
• Increase financial support for the industrial safety administration (BIS) and strengthen law enforcement.

• DeepSeek broke through the limitation of computing power through technological innovation, indicating that it is difficult to curb the development of AI in China simply by relying on hardware control.
• On the contrary, it may prompt China to speed up independent chip research and development and algorithm optimization innovation.

Promote Global AI strategy adjustment

• trump administration announces $500 billion "Star Gate" ( Stargate) plan to invest in AI infrastructure.
• China launched a special comprehensive financial support of 1 trillion yuan, focusing on supporting key enterprises including DeepSeek.
• Other countries are also reassessing the AI strategy and weighing the cooperation with China and the United States.

This event marks that the AI competition between China and the United States has entered a new stage. It is no longer a simple competition of computing power, but an all-round competition of algorithm innovation, business model and ecological construction. The future trend will depend on the game between the two sides in technological innovation and policy adjustment.

3. [international competition] will DeepSeek's success prompt the United States and other countries to adjust their AI development strategies? Will countries increase investment to cope with the new competition pattern?

DeepSeek's success has indeed had a profound impact on the global AI competition pattern, especially in major AI powers such as the United States, which has triggered policy adjustments and strategic reflections. The following is a detailed analysis from three aspects: the U.S. response, investment from other countries and changes in the global competition pattern.

I. Possible strategic adjustments in the United States

1. Increase investment in research and development and technological innovation

• re-evaluation of technical route: DeepSeek's success shows that relying only on large-scale computing and capital investment is not the only way, and algorithm innovation and architecture optimization may be more critical. The United States may focus more on developing efficient algorithms, data optimization, and model architecture innovation.
• Financial support: The U.S. government and enterprises are expected to significantly increase their financial support for AI research and development. For example, large-scale AI infrastructure investment similar to the "Stargate project" may exceed US $500 billion.
• Computing power infrastructure construction: continuously strengthen the construction of high-performance chips and data centers to ensure computing power advantages in AI training.

2. Strengthening export control and security policies

• strengthening chip export control: Although the United States has imposed a strict ban on chip exports to China, DeepSeek's success shows that the effectiveness of such policies is challenged. The United States may further tighten restrictions on key technologies and semiconductor exports to curb China's AI progress.

• Risk assessment of open source technologies: the United States may reassess the management of open source AI technologies, especially open source models in high-risk areas, to prevent technology from spreading to competitors.

3. Talents and education

• talent Training: The United States may train more local AI talents by increasing AI-related university courses and research funding; At the same time, it will simplify the visa process and provide generous treatment to attract the world's top talents.

• International cooperation: despite intensified competition between China and the United States, the United States may strengthen cooperation with its allies in the AI field, such as working with the European Union, Japan and other countries to formulate AI technical standards and ethical norms.

4. Stimulate market competition

• deepSeek's low-cost and efficient model has injected competitive pressure into the US market. U.S. companies may adjust their strategies to meet this challenge. For example, companies such as OpenAI and Meta have promised to launch more powerful models and restart the open source strategy to keep technology ahead.

II. Investment and response from other countries

1. Europe

• policy skew and innovation support: European countries may increase financial support for AI research, especially in algorithm optimization and data governance. Germany and France may jointly promote the AI industry policy at the EU level.
• Supporting local enterprises: Europe may support the rise of local AI enterprises and reduce its dependence on American and Chinese technologies through policies, funds and tax incentives.

• International cooperation: the EU may strengthen cooperation with allies such as the United States and Japan, and strive for more discourse power in AI standard-setting.

2. Other Asian countries

• south Korea: South Korea has quickly responded to the impact of DeepSeek and held a national AI strategy conference. It plans to build a national AI computing center, increase R & D investment and train talents, the goal is to become "AI G3" (the world's top three AI powers) in the future.
• Japan: Japan may focus on the industrial and social applications of AI, especially in autonomous driving, medical AI and other fields, while increasing investment in personnel training and infrastructure construction.

• Singapore: Singapore may continue to increase investment in smart cities and AI governance to attract multinational enterprises to set up AI research and development centers.

3. Other countries

• the rise of emerging markets: emerging market countries, including the Middle East and Southeast Asia, may use deep data resources and policy support to develop AI applications in specific fields. For example, Saudi Arabia plans to invest US $100 billion in technological innovation.

• Breakthroughs in key areas: some countries may concentrate their resources to develop the application of AI in specific fields such as medical treatment, agriculture and education, so as to overtake in corners.

III. New Global Competition Pattern

1. Technical competition

• from computing power to Algorithm Competition: DeepSeek's low-cost and efficient model shows that algorithm innovation and data optimization will become the core of future AI competition. This model may pose challenges to the development strategies that the United States and other countries rely on with high computing power.

• Diversified fields: the competition for AI applications will expand from traditional natural language processing and computer vision to more vertical fields, such as biomedicine and autonomous driving.

2. Talent competition

• Competing for global high-end talents: top AI talents will become the key resources for all countries to compete. The United States, Europe and Asian countries may attract talents through high salaries, scientific research resources and immigration policies.

• Training local technical talents: countries will strengthen education and training and establish a perfect talent training system to reduce dependence on external talents.

3. Competition between standards and ethics

• leading Power of technical standards: countries will compete for the right to formulate AI technical standards to ensure the competitiveness of their enterprises in the global market.

• Ethics and safety norms: With the expansion of AI influence, countries will need to find a balance between technological progress and ethical norms, and strengthen international cooperation to deal with security risks.

4. International cooperation and differentiation

• areas of Cooperation: despite intensified competition, countries may strengthen cooperation in AI ethics, technical security and other fields to formulate global rules and standards.

• Intensified differentiation: in technology and market competition, the international community may have a more obvious differentiation trend. For example, the United States and its allies may form an integrated AI ecosystem, while China may further deepen its cooperation with emerging market countries.

4. [open source ecosystem] How will DeepSeek's open source strategy affect the AI ecosystem? Will this promote technology dissemination, cooperation and innovation, and promote open source models to become the mainstream?

DeepSeek's open source strategy will have a profound impact on the AI ecosystem, mainly in the following aspects:

1. Accelerate technology dissemination

• open-source model weights through MIT licenses, no downstream application restrictions
• significantly reduce the use cost (training cost is about US $6 million, far lower than US $60 million for similar models)

• innovative technologies such as multi-head potential attention (MLA) can be quickly replicated and applied to promote technological progress in the industry.

2. Promote cooperation and innovation

• open technology ecosystem attracts global developers to participate in improvement
• promote deep cooperation between academia and industry and accelerate technological breakthroughs.

• Form an active developer community and focus on intelligence to solve technical problems

3. Open source models become mainstream

• the performance is equivalent to that of the closed-source large model, but the cost is lower and the usage is more flexible.
• Lower the threshold of AI technology to enable more small and medium-sized enterprises to apply AI

• break the monopoly of a few technology giants and promote the sound development of the whole industry.

Through the open source strategy, DeepSeek is reshaping the AI industry structure. This open sharing mode not only accelerates technological innovation, but also makes AI technology truly popular. However, open source also brings challenges such as data security and abuse risks, and corresponding mechanisms need to be established to standardize and guide it.

5. [technological innovation and milestone] has DeepSeek really overturned the AI training mode? Does its technological innovation, such as enhanced learning and multi-modal capabilities, mark a new milestone in AI development?

DeepSeek has indeed brought significant innovations in AI training models, but it may be too early to call it "subversion.

1. Breakthrough in training efficiency

• the training cost is only US $5.57 million and 2048 H800 GPUs are used, which is 90% lower than that of traditional large models.
• Through the optimization of FP8 precision, modular architecture and proprietary technologies such as DualPipe, the training efficiency is improved qualitatively.

• In terms of inference costs, multi-head potential attention (MLA) technology significantly reduces KV cache requirements.

2. Core technology innovation

• the first large-scale application of multi-Token prediction (MTP) technology
• hybrid expert model architecture to implement accurate Token routing through gating network

• the R1 model improves the reasoning ability by generating synthetic data and enhancing learning, and approaches or surpasses the GPT-4 in tasks such as mathematics and code.

3. Market impact assessment

• this leads to fluctuations in Nvidia's share price, reflecting the market's expectation of a decrease in GPU dependence.
• However, the total demand for GPU is expected to continue to rise, as lowering the threshold will bring more market participants.
• It is estimated that by the end of 2025, the training cost of similar performance models is expected to be reduced by another 5 times.

6. [AI autonomous reasoning ability] what does DeepSeek's "epiphany moment" phenomenon mean? Does AI have a higher level of autonomous reasoning capability?

What does DeepSeek's "moment of epiphany" mean?

1. Signs of technological breakthrough:

• self-optimization ability: the model can re-evaluate the problem-solving ideas during training, actively correct errors and try more efficient strategies. This is similar to the "flash of light" of human beings, that is, to discover problems instantly and find solutions.

• Inference chain optimization: The model automatically generates a longer and more accurate inference chain ( CoT). For example, in complex mathematical problems, the problem-solving process of the model changes from a simple method to a more complex but correct path, showing the autonomous adjustment ability of the reasoning process.

2. The potential of reinforcement learning:

• completely dependent on reinforcement learning training: DeepSeek-R1-Zero tests the deep application of reinforcement learning in complex tasks, does not rely on traditional supervision and fine tuning, and directly optimizes reasoning methods through reward mechanism. This method not only reduces the dependence on large-scale annotation data, but also improves the adaptability and reasoning ability of the model.

• Significant performance improvement: In the AIME 2024 benchmark test, the pass @ 1 score of DeepSeek-R1-Zero increased from 15.6% to 71.0%, and reached 86.7% after the majority vote. This non-linear breakthrough demonstrates the great role of reinforcement learning in unlocking AI potential capabilities.

3. Inspiration from innovative methods:

• deepSeek's research shows that competitive AI models can be developed even under the condition of limited computing resources through technological innovation. This provides a new direction for future AI research and development: instead of simply pursuing larger models and more data, it is better to focus on training methods and algorithm optimization.

4. Impact of global AI competition:

• "Nik moment" Alert: DeepSeek's success is compared to the "artificial satellite moment" in the AI industry, which reminds the United States and other leading AI countries that AI competition not only depends on hardware resources, it is also closely related to algorithm innovation.

• Inspiration to policies: the United States and the world need to re-examine the investment methods of AI research and development, especially the investment in strengthening learning and reasoning capabilities, in order to prevent being surpassed by technological innovation.

Does AI have a higher level of autonomous reasoning capability?

1. Significant Improvement in reasoning ability:

• self-verification and reflection: The model can discover its own problems and actively adjust them in the reasoning process. For example, in a code generation task, the model verifies whether the initial scheme is feasible. If it fails, it attempts another path.
• Chain thinking ability: through the thinking chain (CoT), the model can decompose complex problems into multiple steps to solve one by one. This method significantly improves the performance of the model in logical tasks such as mathematics and physics.

• Problem-solving efficiency: when facing more complex tasks, the model will automatically extend the inference chain, which is similar to the behavior that human beings spend more time thinking when facing difficult problems.

2. Limitations of Autonomous reasoning:

• depending on Setting task objectives: the current AI can only make reasoning within the preset task scope, and its autonomy is limited to the execution level, not the target setting level. For example, a model can solve mathematical problems, but it cannot decide the goal of "solving mathematical problems" by itself.

• Lack of real cognitive ability: AI reasoning is essentially based on data-driven pattern recognition, rather than awareness or abstract cognition. Its behavior is more like the algorithm output optimized by reward mechanism than the reasoning in human sense.

3. Future potential:

• evolution from task-specific to general intelligence: DeepSeek's success shows that AI's reasoning ability in specific fields has approached or even surpassed human beings. However, to realize general intelligence (AGI), breakthroughs still need to be made in cross-domain adaptability and goal setting ability.
• Infrastructure investment: higher-level autonomous reasoning requires more powerful computing resources and more advanced algorithms. The success of DeepSeek also shows the potential of reinforcement learning. In the future, more optimized training methods may be used to achieve reasoning capabilities closer to human beings.

The above content is transferred from Andy730