Chinese AI firm DeepSeek has taken the tech world by storm with its LLM and reasoning model, sparking news and speculation. As it challenges OpenAI and the US, the truth is unfolding. Whats real, whats not, and why does it matter? Let’s break it down
What is DeepSeek?
DeepSeek is a Chinese AI firm backed by High-Flyer, a quantitative hedge fund based in China. Liang Wenfeng founded it and operates independently. The firm recently released an AI reasoning model that claims to have outperformed OpenAI’s ChatGPT o1 on several AI benchmarks and has sparked conversations, controversy, and admiration around the globe. DeepSeek R1, an open-source reasoning model, and DeepSeek-V3, a large language model (LLM), were released in January 2025. The DeepSeek mobile app, which provides a chatbot interface for R1, leaped to the top of the Apple App Store only a few days after its release, making it more popular than OpenAI’s ChatGPT in the US and the U.K.
You may also be interested in Mastering Strategic Adoption of AI: Your Roadmap to Success
How does R1 work and how close is it to OpenAI’s o1?
R1 is designed to perform reasoning, problem-solving, and language-based tasks efficiently. It is built on a transformer architecture that processes text in parallel, enabling fast and
efficient language modeling. Like OpenAI’s GPT models, the model has been trained on an extensive collection of books, papers, code repositories, and even data from some of OpenAI’s models. R1 also features chain-of-thought reasoning, instruction-
tuning, and reinforcement learning from human feedback to improve the quality and relevancy of its responses over time.
Regarding reasoning and problem-solving, benchmarks published by DeepSeek indicate that R1 is comparable to o1 in general mathematical datasets, but o1 has an edge in complex, multi-layered problems. For coding, R1 is capable of generating, debugging, and optimizing code. Its performance is similar to GPT-4-turbo in languages like Python,
JavaScript, and C++. On the other hand, o1 may have a slight edge in solving advanced programming challenges. However, it’s essential to remember that the models are not identical and do not necessarily lend well to strict comparisons.
The importance of DeepSeek innovations
The R1 model has demonstrated exceptional performance, particularly in tasks involving mathematics and coding, which is incredible for a company that was barely noticed in China’s AI community until 2024. Initially, it was reported that the model was trained with significantly fewer resources, making it much less expensive than most other models; however, that may be far from the truth. According to a report by SemiAnalysis, DeepSeek’s 6 million dollar figure only applies to pre-training, which is a tiny portion of the development of the entire model.
They estimate the figure to be more like $1.6 billion, which includes R&D and hardware, which is estimated to be more than $500 million alone. However, just because the innovation’s overall price went up does not make it any less impressive for a Chinese firm. SemiAnalysis compares R1’s overall costs to other models, noting that as AI development continues, DeepSeek has become unique in that it has been able to capitalize on previous advancements amid strict restrictions to achieve a new level of costs relative to capabilities.
The current state of affairs and controversy
Market Disruption
Global tech stocks had a significant sell-off following the launch, with the Nasdaq Composite dropping 3.4%. Nvidia shares fell by 17%, resulting in a loss of nearly $600 billion in market value. Other tech giants like Google and Microsoft also experienced declines.
Global Reactions
The success of DeepSeek models has been described as a Sputnik moment for AI by tech investor Marc Andreessen, highlighting China’s rapid advancements in the field. The development has garnered the attention of the United States and other governments, considering the US’s strict regulations surrounding access to advanced chips drive a company’s ability to innovate by providing the computing power needed for AI. While this seems to have necessitated more
innovative approaches to building AI pave the way for tech companies outside the US to do more with less.
Large-Scale Malicious Attacks
Shortly after its release, DeepSeek servers were hit with more than 230 million distributed denial of service (DDoS) requests per second in the first 83 hours.
You may also be interested in Safeguarding Your AI Systems: How to Prevent Prompt Injection Attacks in LLMs
What about DeepSeek censorship and security?
DeepSeek’s AI has faced criticism for strict censorship on sensitive topics and concerns about data privacy. It stores user data on servers in China and is subject to China’s laws around sharing user data with authorities. When using the company’s proprietary app or service, questions about China’s sensitive history have been met with blank responses or Beijing’s canned propaganda. However, if the model is run on a different tool without the cloud
firewalls and protections, the model will return a more complex and involved answer. As with any innovations coming out of China, the conversation around censorship will be nuanced and is not likely to be resolved soon. On the other hand, the fact that the model is open-source and can be used on US-based or private servers, making it more attractive to some users.
What’s next for reasoning models like R1?
Reasoning models will continue to become more prolific, cheaper to develop, and better at surpassing previous benchmarks than their predecessors. Google released Gemini Flash 2.0 Thinking a month before DeepSeek’s R1, which coincidentally outperformed R1 in three reported benchmarks.
DeepSeek’s recent releases have sparked significant discussions about the future of AI development, international competition, and the balance between resource investment and technological advancement. They have ushered in a new era of innovation amid constraints. As we enter another era of AI advancements, the story is just beginning to unfold for DeepSeek, NVIDIA, OpenAI, and various actors.
