DeepSeek’s arrival on the AI scene has sent stocks of US-based global technology companies tumbling, raising questions about their valuations and the position of the US as the technology leader in this area.  

Founded in 2023, DeepSeek works to develop open-sourced AI models. Its founder, Liang Wenfeng, runs a quantitative hedge fund, High Flyer, which has been known for using advanced computing to analyze financial data since 2015.?

Liang’s scientific curiosity soon took over, and he created DeepSeek to develop groundbreaking models that could be inspected and improved by the developer community and not aimed at generating monetary returns. 

DeepSeek’s mobile app was released in January this year and has surged to the top of the iPhone download charts. Unlike OpenAI’s ChatGPT, DeepSeek’s chatbot articulates its reasoning before giving a response to a prompt, Bloomberg reported.

However, DeepSeek’s advanced reasoning model R1 has shaken the ground for the US tech sector. Although the details have not been revealed, the cost of training and developing the model is only a fraction of what OpenAI or Meta spent creating their large language models (LLM).?

While companies have deployed supervised fine-tuning to develop their models, DeepSeek claims it has used reinforcement learning (RL) to achieve robust reasoning abilities. The Bloomberg report added that its latest model, R1, matches the performance of OpenAI’s 01 models in reasoning tasks and trumps rival models on benchmarks for mathematical tasks, general knowledge, and question-and-answer performance.?

DeepSeek has achieved all this using a fraction of the computing power used by Meta to develop their LLMs, while the company’s registered capital is just 10 million yuan ($1.4 million). This is quite against the trend of AI companies, even those based in China seeking more funding to improve their models in the future.?

To avoid a scenario where China becomes the technology leader in AI, the US imposed export controls on computing hardware such as NVIDIA’s chips as early as October 2022. DeepSeek’s journey began with 10,000 H100 chips, but the export controls meant it could not compete with US companies using the same approach, according to a Business Standard report.?

Unlike other Chinese AI firms that took the application development approach, the team at DeepSeek worked on reimagining the underlying architecture of AI and optimizing resource efficiency. This included custom communication schemes for efficient data exchange, memory optimization, and combining smaller models to achieve superior results. 

Apart from its flagship R1 model, the company has six distilled smaller versions ranging from 1.5 billion to 70 billion parameters. These are MIT-licensed, free to use, and allow researchers to fine-tune and commercialize their work.?

This will push companies like OpenAI and others to lower their pricing to maintain their lead but question their operating models in the face of highly efficient ways of deploying AI. This has spooked investors, triggering a huge sell-off of tech stocks in the market. 

We may be witnessing a sea change in how AI will look in the future.