DeepSeek has taken the internet by storm.
This week, the Chinese AI company DeepSeek gained massive popularity as its chatbot application skyrocketed to the top of the Apple App Store and Google Play charts. The company’s innovative AI models, which utilize compute-efficient methodologies, have prompted analysts on Wall Street and tech experts to reassess whether the U.S. can hold onto its leadership position in AI and if the demand for AI chips will persist.
But what is the backstory of DeepSeek, and how did it gain international recognition so rapidly?
The Trading Roots of DeepSeek
DeepSeek is funded by High-Flyer Capital Management, a Chinese quantitative hedge fund that leverages AI in its trading operations.
Liang Wenfeng, an AI enthusiast, co-founded High-Flyer in 2015. He initially ventured into trading while studying at Zhejiang University and established High-Flyer Capital Management as a hedge fund in 2019, focusing on crafting and implementing AI algorithms.
In 2023, High-Flyer launched DeepSeek as a research lab focusing on AI tools, separate from its financial pursuits. With backing from High-Flyer, DeepSeek emerged as an independent entity.
From its inception, DeepSeek constructed its own data center clusters for training models. However, like many AI firms in China, it faced challenges due to U.S. export restrictions on hardware. To develop one of its latest models, the company had to resort to using Nvidia H800 chips, which are less powerful than the H100 chips accessible to U.S. firms.
The technical team at DeepSeek is reported to be quite young. The company actively seeks out PhD-level AI researchers from prominent Chinese universities. Additionally, DeepSeek employs individuals without computer science backgrounds to broaden its technological understanding across various fields, according to The New York Times.
Robust Models from DeepSeek
DeepSeek launched its initial suite of models—DeepSeek Coder, DeepSeek LLM, and DeepSeek Chat—in November 2023. However, it wasn’t until the following spring, with the introduction of the next-generation DeepSeek-V2 model family, that the AI sector began to take significant notice.
DeepSeek-V2, a versatile system for text and image analysis, achieved remarkable results on several AI benchmarks while being notably more cost-effective than similar models at that time. This compelled domestic competitors, including ByteDance and Alibaba, to lower prices for some of their offerings and make others completely free.
The release of DeepSeek-V3 in December 2024 further cemented DeepSeek’s reputation.
As per DeepSeek’s internal benchmark evaluations, the DeepSeek V3 outshines both downloadable, open-source models like Meta’s Llama and “closed” models only accessible via API, such as OpenAI’s GPT-4.
Impressively, DeepSeek’s R1 “reasoning” model, unveiled in January, is claimed to perform comparably to OpenAI’s o1 model on critical benchmarks.
As a reasoning model, R1 effectively self-verifies its outputs, helping it navigate common pitfalls encountered by typical models. Though it may take slightly more time—ranging from seconds to minutes—to produce results, it generally demonstrates higher reliability in areas such as physics, science, and mathematics.
However, there are drawbacks to R1, DeepSeek V3, and the company’s other models. Being developed in China, they must adhere to regulatory benchmarks set by China’s internet authorities, ensuring that responses align with “core socialist values.” For instance, R1 refrains from addressing inquiries about Tiananmen Square or Taiwan’s sovereignty within DeepSeek’s chatbot app.
A Game-Changing Strategy
While DeepSeek appears to have a business model, the specifics remain ambiguous. The company offers its products and services at prices significantly lower than market norms and even provides some for free.
According to DeepSeek, innovations in efficiency have empowered it to maintain exceptional cost competitiveness. However, some analysts challenge the credibility of the claims made by the company.
Regardless of varying opinions, developers are gravitating toward DeepSeek’s models, which, while not open-source in the traditional sense, come with permissive licenses that permit commercial use. Clem Delangue, CEO of Hugging Face, noted that over 500 “derivative” models of R1 have been created by developers on Hugging Face, collectively amassing 2.5 million downloads.
DeepSeek’s achievements in outpacing larger, more entrenched competitors have been characterized as “disruptive” and “over-hyped.” Notably, the company’s success contributed to an 18% drop in Nvidia’s stock price in January, and prompted a public commentary from OpenAI CEO Sam Altman.
Microsoft announced the integration of DeepSeek within its Azure AI Foundry service, consolidating AI offerings for enterprises under a unified platform. During February’s earnings call, when asked about DeepSeek’s influence on Meta’s AI expenditures, CEO Mark Zuckerberg acknowledged that ongoing investment in AI infrastructure remains a “strategic advantage” for Meta.
In Nvidia’s fourth-quarter earnings call, CEO Jensen Huang highlighted DeepSeek’s impressive innovation, asserting that it, along with other “reasoning” models, presents significant computation requirements for Nvidia.
Conversely, some corporations and even entire nations, including South Korea, have prohibited the use of DeepSeek. The state of New York has also restricted DeepSeek from government devices.
As for DeepSeek’s future trajectory, it’s uncertain. Enhanced models are anticipated. However, the U.S. government seems to be growing increasingly cautious about perceived foreign threats. In March, The Wall Street Journal reported that the U.S. is likely to impose a ban on DeepSeek for government use.
This article was initially published on January 28, 2025, and will receive regular updates.
Compiled by Techarena.au.
Fanpage: TechArena.au
Watch more about AI – Artificial Intelligence
